I'm getting fatigued with the constant introduction of names and the impenetrability of mapping these brand new names to use cases.
I think internally, Google, Anthropic and OpenAI are thinking "It's so easy, you just use Neon for X, Carbon for Y" - but most of the world is trying to map Sol, Astra, Neon, Opus, Fable, Mythos, Carbon, Haiku, etc.
I just hope they finally land on a naming scheme they are happy with and just stick with it. But I have my doubts. Perhaps they are all benefiting from this confusion.
>m getting fatigued with the constant introduction of names and the impenetrability of mapping these brand new names to use cases.
Because probably there is no system. They just train a model, and find out that it's "kinda good", and that's it, they publish it. The models do not constitute even a partially ordered set, nobody can predict their ranking for your particular tasks.
True story about Sony naming: I was at the PC industry trade show where Sony launched the new VAIO computer in a big, beautiful booth at the front of the hall. Being a long-time Sony fan, near the end of the first day after the crowds had mostly cleared, I stopped by the Sony booth on my way out to see the new PC. I saw that the VAIO name also formed the computer's distinctive ultra-modern logo featured in large bold text on each side of the PC's futuristic case in a minimalist block font. But I was puzzled how a company with a history of so many great products with awesome names (WalkMan, DiscMan, PlayStation) had come up with such an awful name.
Since I was curious about the origin of the name, I wandered over to where three Sony employees were chatting passing time until the day ended. I learned all three were mid-level marketing staff based in Sony's New Jersey US HQ. So I asked "How do you pronounce the computer's name?" All three had different takes between various long vowel vs short vowel sounds.
Since they were all facing me, I also asked, "Without turning around, how do you spell the name?" They all paused, then laughed. They did manage to work that out after a few false starts but it wasn't exactly smooth. As we chatted some more I asked who'd come up with such a unique name and learned Sony had hired a "really cool" high-end branding agency who specializes in coming up with great product names. The effort and expense in creating the name and logo had clearly been marketed to employees as "cutting-edge cool." I warmly wished them the best with the new product before wandering out of the hall to dinner.
None of it true of course. The VAIO name and logo was by Gotō Teiyū who had been working as a designer at Sony for 30 years at that point. Its pronounciation is quite unequivocal in Japanese as well (like bio).
Well, I guess the mid-level marketing people in Sony's New Jersey office got some garbled version of what happened in Japan, because that's what they told me at PC Expo '96 at the Javitz Center.
Regardless, the PC and it's neat industrial design plus Sony's brand and distribution were good enough to overcome a weak name. VAIO was an acronym for "Video Audio Input Output" (which Sony later changed to "Visual Audio Intelligence Organizer", although I'm not sure why they thought that would be better).
There is definitely a system. They are training models of different sizes for different levels of inference efficiency, which maps to token pricing. The ordered set is the size of the weights/ computational complexity during inference.
I like both naming schemes...but the thing that always confuses me is, aren't fables pretty short? The Tortoise and the Hare is a lot shorter than say, Anna Karenina or other opuses...
Although, maybe it's called Fable because it's trying to bluntly convey strong moral lessons, hence the strong safety classifiers...
I like them too but it took a while for me to land, tbh. Mistral does the best job IMO: Mistral Small/Medium/Large. Slap a number on for the generation.
It makes sense if you understand that the size of a model is something important and meaningfully different than just how advanced the model is.
You could argue that's a detail the user shouldn't have to concern themselves with, but I think it's appropriate to expect them to learn that and pick names which put that concept front and center. The various "model router" approaches for consumer tools like ChatGPT have been finicky and inappropriate for professional tools we spend a lot of money on.
At work at Giant Tech Company, someone has started publishing a "WTF" guide every week that includes the decoder ring for model names vs. their capabilities. I just have my agent read it every week and update all my skills and rules appropriately to figure out which model to use for what.
After looking over the landing page, my first thought was also about whether I wanted to invest the time to figure out where this new model fit in the relative hierarchies of comparable product stacks. Ultimately I decided "Nah", and aborted to the comments.
This sounds highly likely.
Imagine trying to name and market a bag of weights, and arguing with a straight face why it’s better than another bag of weights.
I maintain a table for all labs. each row is S / M / L / XL (maybe 2XL 3XL etc in future) and each column is the lab. The only way i can contextualise any of this.
They really need an industry term for each model. Think like mobile, tablet, laptop, desktop. Apple, Lenovo, HP, Google, and others all have their own versions of these, but in general the public has some idea of common use cases that each product will satisfy without delving into the details very much.
to be fair we have certain semi-common names like "flash" and "small", even though those tend to be relative to the size of their flagship model, which makes the terms somewhat meaningless out of context... but at the very least those do help when you're familiar with the given company.
Yeah when I talk to people who don’t follow ai closely they are very confused, particularly about Claude where you have to pick a model by an arbitrary name. At least ChatGPT has a slider for model strength
Anthropic has at least been consistent these other ones change it every fucking week. I honestly just think of models in terms of haiku/sonnet/opus level now because the other ones are completely inconsistent.
There's no "Fable model". There's Mythos. And it's certainly larger than Opus.
Fable is Mythos with safeguards, basically.
In fact, Mythos also includes safeguards, so at this point it's not really clear what's the point. You can apply for security program and there are three levels of access to the Mythos. I guess now it makes sense to just drop "Fable" name and go with Mythos which will just be implicitly restricted according to your access level.
Or maybe they'll realize that there are abliterated openweights models, and restricting Anthropic models does more harm than good. I literally have to use other service for security audit of my software, because Fable rejected it.
I'm terrible with names, so I'm in the same boat. There's a lot of marketing going on.
But the naming isn't completely random, here's what I use to make the ordered list in my head:
Claude is named after creative works, and uses bigger works for more powerful models. In order of strength:
* Haiku - short poem (we all know this one)
* Sonnet - "medium" length poem (many of us know this one)
* Opus - It's the singular form of "opera", so it's like a stand-alone segment of a larger body of work. (Nobody knew this one, but not hard to remember, and know you know that "opera" is actually plural! The description sounds bigger than a sonnet, but it's not a strict ordering.)
* Fable - a story (this one doesn't seem obviously bigger than an opus, but I guess an opus is one part and a fable is a front-to-end story, it works)
* Mythos - another type of story, which works and it's basically Fable but Anthropic's internal-only version
Gemini, as of version 4, started doing something with atoms:
* Hydrogen - smallest periodic element, the most common gas in outer space
* Neon - a bigger element, less common in outer space
* Argon - yet bigger element, yet less common in outer space (as best I can tell, I am not an astronomer)
* Carbon - this breaks that pattern, its a smaller element than the first three, and not a gas... but critical for life on this planet? And it sounds like a longer version of Argon. So I can remember it, but I'm not sure what the intended pattern was.
The model ordering is not strictly upgrades, but AFAIK they are intended to be steps up in terms of (claimed) complexity and reasoning power. But since it's an evolving field, the specifics of how cost, speed, and accuracy combine can make it hard to know how to order a given model for your use case. (i.e. cheap & fast with a harness that makes more calls, or expensive and slow but better at one-shot).
Deepseek's naming is more organized and literal, but you still have to remember the prefixes for V = version (plain) and R = Reasoning, and that they have different specialized models that aren't intended to be compared to each other.
Basically every lab has a model that doesn't fit their naming scheme 100%.
OpenAI's naming scheme is in Latin:
* Luna - Moon (smallest)
* Terra - Earth (medium)
* Sol - Sun (large)
* Astra - Star (could be larger than the sun I guess)
Meta's Muse:
* Ember - (smallest)
* Spark - (medium)
* Wonder - (large)
A spark is the beginning of an idea whereas wonder is associated deep thought. But an ember and a spark are both small pieces of fire, so the ordering is a little muddy.
Google's naming scheme has the least clear pattern of any lab.
I wonder if it was supposed to grow in fire size. Ember and spark fit the pattern and there’s plenty of room to grow. But eventually they’d start reaching sizes related to catastrophes.
Speaking of Opus, many have likely heard of the expression "magnum opus" (Latin for great work). That's probably easier to remember than the singular form of opera.
Seriously. It's annoying enough to keep track of the names of which models mean what just within Claude Code. But they don't change that often, so I can deal.
But trying to keep track of the model names in all 3 companies? Utter madness.
We need something better. All I really care about most of the time is whether I'm using the cheapo, average, or expensive model. I don't really care about which actual model is occupying which actual spot at which time with which cutesy name. I just know I have a budget, and so I allocate the three levels according to the type of work I'm doing.
I think vendors assume their names schemes are obvious.
Anthropic: "We'll name after the size of literature style! An opus is obviously bigger than a sonnet!"
Much of the world: "What is this word Opus?"
OpenAI: "We'll use the Latin names for astronomical bodies. Everyone knows that terra = earth and luna = the moon and that the earth is bigger than the moon."
Much of the world: "Does terra mean it is terrified? Oh well at least it is better than Gpt 4o vs Gpt o3"
Google: "We'll name it Bard! Oh, that was bad. Lets name it Gemini because that sounds cool. Oh we need different models, ok we'll have Pro for the big expensive on, and Flash for the fast one. But now we want a fast large one and also we have mid sized models. Ok, lets name them after elements, but not ones that anyone knows anything about so now we can confuse them as to if Neon or Carbon or Argon is better. But Gemini Hydrogen is the cheap fast one"
Everyone: "Ok, Gemini Hydrogen is cheap. But this is too confusing, I'll just use Claude".
It doesn't help that the blog post opens with a long description of all the benchmarks that the model excels on, but very little info about how I should pick what model to use.
I do note that gemini.google.com has replaced Argon with Neon, so I guess maybe that's a clue?
My personal approach is to not really care at all. These specific models aren't going to be here in 6 months, so learning the intricates of these specific models has no benefit for me. I'm just going to mostly rely on whatever one is the default
It's possible they are just making it confusing as possible so everyone switches to the orchestrator but I bet they are just as confused as the rest of us.
This wouldn't be a conversation if the big 3 didn't change their model names multiple times per year. Or conversely - do you think people would be fine with Toyota renaming the Camry every 6-9 months?
Maybe the trick to working with these RaNd names is for us to tell Claude Fable/Opus or GPT Astra/Sol: "go find the Gemini model that is as smart as you and discuss XYZ with it, and then tell me what you guys agree on."
I feel like Google purposely names their models not for humans but for AI agents. People keep saying: "make your products ready for agents," so maybe that's what Google is doing, which is making Gemini 4 models ready for other agents to use.
We've moved to <family> <variant> <number> e.g. deepseek flash 4.1. I can keep up with that, and if/when Flash 4.2 comes out, great. I can grok that. Google's naming convention just shows how out of touch management is with the market.
Android's silly food naming convention works when you do one release every ~18 months and you are the market leader. It makes zero sense when you're releasing 8 models a year
Clicked this expecting the top comment to be a gripe about naming conventions, was not disappointed. But the trope is getting a little old at this point. It might be time to consider another irrelevant talking point you can beat to death.
I dont agree that Anthropics names are easy for an outsider, i'd have no clue what a Fable is compared to Haiku, to me both are just "story" names. OpenAI's new naming scheme is celestial bodies and i'd assume it's just size, bigger celestial body bigger performance, except i dont know where "Astra" fits since Sol > Terra > Luna at least make sense. Now gemini... I was just thinking it today that "Carbon" and "Hydrogen" with hydrogen being the lightest element? Is that what they are going for? I wouldn't know off the top of my head where Neon and Argon are placed though.
Edit: Just as I post this someone else comments google's doesnt follow elemental sizes at all so my theory is wrong too!
I'm obviously thick, but this is the first time it's dawned on me that the Claude names were types of written work. Now I'm wondering why Opus isn't the top model; isn't an opus a more impressive (or at least larger) piece of work than a fable?
I don't have a cohesive explanation for each one, but in my mind a fable is basically a legend (or in other words, a myth/mythos) and a legend is the final form of a story. Even though it has been half forgotten and changed, it still lives on, affecting people across times greater than a human's short life.
A fable is just a story with antropomorphized animals and a moral. They can get very short.
Here's a translation of the entire Fox and Grapes fable:
> Driven by hunger, a fox tried to reach some grapes hanging high on the vine but was unable to, although he leaped with all his strength. As he went away, the fox remarked "Oh, you aren't even ripe yet! I don't need any sour grapes." People who speak disparagingly of things that they cannot attain would do well to apply this story to themselves.
An opus is just a musical composition, sometimes also used about written work, but it doesn't really imply a great work. Maybe you're thinking of "magnum opus"?
The point is that Haiku was Haiku a year ago, and the number keeps increasing, while googles gemnini started with sane names and numbers as 1,2,3.x pro/flash/lite, and is now suddenly continuing with 4 hydrogen/argon/neon/carbon something.
For me, it is familiarity that does the work. I'm familiar with Sonnet, Opus and Fable and what they mean to me. The same is true for Sol and Astra bu I'm not sure about Terra or Luna.
So I'm just hoping for consistency over some period of time to allow me to get familiar.
Prior knowledge of common poetic forms taught in middle school?
The point is that there is prior knowledge that completely explains the entire system. A 3 line poem < a 14 line poem < a book-sized work < a multi-book epic, boom done, easy. Consistent and friendly using words that everyone should know.
With google we have element 1 < element 18 < element 10 < element 6. There is no consistent schema where their namesakes map to their relative sizes, price, and capability.
They still require prior knowledge of Latin words of celestial bodies or different words for literary works, same as Google's requires knowledge of the periodic table.
Damn, if only the people concerned with what the naming of AI models mean had access to some magical technology they could use to ask what the names of those models mean.
I mean, they're OK, still a bit cryptic. Astra in particular is a bit ambiguous - they're all named after celestial bodies then the largest is just named after space in general.
Would love "small", "medium", "large", "extra-large", etc...
Having named a range of sizes more than few times by now...S/M/L/XL always starts out great and then turns into a MESS of XS, XXXL, Medium-Large, forever and ever.
Yeah, or you get an escalatory scale. Similar to how LF/MF/HF (low/medium/high frequency) got "very HF", "ultra HF", "super HF", "extremely HF", "tremendously HF".
Google used to do that with a version number and descriptor. 2.5 Pro, 3.1 Flash-Lite, etc. Gave you both model generation + idea of relative size/quantization within generation. I guess they decided that scheme wasn't flashy enough?
I don't think it's just due to AI; the toxicity and cultist behavior of the active core HN users has been annoying for years; even Karpathy complained about it openly on X a while ago.
Is there any theory on what is Anthropic's secret sauce? I have been using the latest Gemini models and they just don't seem as good as Opus 5.5 despite the benchmarks. Yeah it comparable on some tasks maybe slightly better but as an overall general intelligent model nothing compared to Opus 5.5 and it hard for me to understand why. Opus feels like human level intelligent like I'm talking to a coworker basically while GPT and Gemini feel like I'm talking to an AI and have to change how I talk. It hard to explain and maybe I'm the only one that feels this.
Is it because Anthropic put a ton more effort into the character of the model that it basically roleplays as a human and that why it tends to think better when giving tasks that it might not be trained heavily on.
I have used Carbon more than most model to try and get the same results but it just feels like i'm talking to a smart AI. Which is fine if you are asking questions, researching an idea, but if you actually want to work on something, build something, it just falls apart. Like Opus can run for days on a task and it feels like it understand what it was doing 2 days ago even when it might not be in it context while other models feel like they breakdown and get lost or falter as the task grows.
I think Gemini 4 (Argon is the one I've been using) is very clearly better at writing English than any Anthropic model. It's not better otherwise but in that the difference is vast.
To the point that I now use a workflow of delegating from Opus or Fable to Gemini4 to write commit messages and PR descriptions, it mostly solves the Claudese problem.
> To the point that I now use a workflow of delegating from Opus or Fable to Gemini4 to write commit messages and PR descriptions, it mostly solves the Claudese problem.
how do you do this w/o breaking bank? for Argon to write good commit messages, it must read the diffs and all the work Opus has done, which could easily eat up its usage quota.
I use a $20 a month plan with Argon and I honestly find it pretty usable for asking a lot of questions if I don't keep it all in one session, or it can do a few small coding tasks a day for me until I hit my limit. I haven't hit any limits with the smaller models cuz I don't like using a lot of LLM code in my personal projects since me learning a new thing or area by coding is one of the main motivations. Generating messages from diffs/notes after the work is done is pretty cheap I imagine.
Gemini has reliably been the best writer of the bunch, and Argon has blown me away with how "human" it can sound. It has so much less of the AI smell than the others, though Opus 5.5 definitely lobotomized the absolute sloppery that was Opus 5, at least.
Every day I'm just more and more amazed at Opus 5.5 and how ahead it was. Even with the insane pace we're going at, it's amazing that two months later, we're barely putting out models that hit the price/speed/quality of Opus 5.5
No wonder they went for the 5.5 naming, instead of the .1 increase. From just the benchmarks, it seems like neon is basically on par with Opus 5.5 which came out 2 months ago.
Same here. Spin up a few long horizon tasks, have it manage the whole thing as a project, you can talk to it, interrupt it, engage in side-quests, its like talking to a human.
Have another session on another machine (say, IOS build on a mac, and server build on a linux?). No problem, they will happily coordinate with each other through the /remote-control protocol, and then you feel like you are in a crowd, with those sessions talking to each other and over you.
No such feel with OpenAI nor Google models for me. They are absolutely great for lots of things, but nothing comes close to the feel of Opus 5.5 and Fable.
(I am biased; i mainly use the anthropic models so i got used to them).
The first time I tried to use fable 5.2 it made a bunch of really amateurish technical design assumptions about my project that cursory reading of docs would have disproven (but also should have just been discarded because of taste reasons anyway). I switched back to Opus 5.5 after that.
I agree Opus 5.5 is something different entirely. It is even hard to find tasks for other agents to be honest where you know the results will be picked apart when Opus 5.5 reviews them. In my setup Opus 5.5 uses codex from time to time when I'm running out of credits and it decided to run 2 astra agents for the same task to improve its performance.
Anthropic spent a lot more time editing their pre-training data with higher-quality material. For example, they used actual books for pre-training, because books are generally run through human editors for higher quality. (this is from the copyright lawsuits)
Other companies would use completely raw internet data for pre-training, which is basically trash. Anthropic's secret sauce is that they filtered and edited down their data for only the highest-quality material. You really don't need quadrillions of tokens to pre-train, as that's just random noise. A smaller-but-higher-quality data set is better in every way.
Meanwhile its obvious that companies like Google would use trash data like email and chat logs or Internet forums for pre-training. Just complete trash data.
This is a classic case of "less is more" and "garbage-in-equals-garbage-out"
Anthropic spent a lot more time editing their pre-training data with higher-quality material. For example, they used actual books for pre-training, because books are generally run through human editors for higher quality. (this is from the copyright lawsuits)
[ ... ]
Meanwhile its obvious that companies like Google would use trash data like email and chat logs or Internet forums for pre-training. Just complete trash data.
Google has been scanning every book they can get their hands on for 20 years or more. They have a bunch of patents on non-destructive book scanning gear. Surely they are using that data to train their models.
I don't think it pre-training, because surely these labs all have similar data and Google should be very good at ingesting, cleaning, and filtering data.
I honestly think it just their post-training is fantastic and I think it has something to do with them being very opinionated on how they want the model to act that it basically the best at roleplaying a human. Humans are very good at completely and figuring out things that there aren't "trained" on if given basic set of intelligence. So when it working on these long tasks that might not be directly in it training data it still reason through it very well.
They got made fun of for this in the pass because they do a lot of research on this including hiring a philosopher.
This is just my best guess, because you don't see any other labs talk about how they made the model act and Anthropic is basically obsessed with it and talks about it every model release.
I think they just have better "taste" and that taste is their secret sauce.
Microsoft (MSAI) showed that quality >> quantity with their Phi model a couple of years ago. It was a small model, but trained on very high quality data, and did better than much larger models.
I can't say what precisely, but other big companies are also doing what they can to get human data of high quality, just look for all the companies that popped like mushrooms that do ai reinforcement tasks, micro1, ethos, mercor ecc
Anthropic doesn't want people to use their models for the chat/persona style use cases OpenAI has spent more time and effort optimizing for.
If you look back at GPT2 and the early days of AI safety/risk this was actually something the kinds of AI researchers that joined Anthropic didn't want - RL on coding->using that accelerate development and to bootstrap capabilities across domains was always the plan, not making it easier for people to spam/fake homework or trap people in problematic human-AI attention/relationship tarpits.
I personally find that Opus 5.5 and Fable 5+ are generally fine to chat with but just very wordy and culturally "LessWrong/essay-guy". Which, if a model is going to have a personality, is close to the one that I want (though it does piss me off when it gets preachy about standards Anthropic clearly doesn't hold themselves against)
I also want to immediately recognize AI generated content on the web or from another person, so I'm very glad it has a distinct voice
Yeah maybe that's the case, and maybe I am more sensitive than other people. But I don't really use AI for that sort of thing but even for technical stuff I still find it's writing style a real turn off. Gemini still has lots of AIisms but it's a lot more tolerable.
If the training set is smaller, doesn't that make sense though? There will be much less variety in its style and so its going to make very recognizable output.
Yeah that could be it. I guess the thing about what makes AI writing so repulsive sometimes is not that it's bad like a human would be bad, but that it does the same things over and over.
Purely outside looking in, but yes, I'd argue you are right and higher-quality is a major part of it, both for pre- and post-training. There appears to be a far greater willingness to adopt specifically tailored training sets from third-parties, consistently update and build on them.
From simple frontend work to strategically adding virtual rulers when working with LIDAR scans to extract more accurate data, there are very clear points when Anthropic introduces a new external data set into their mix and while I have seen the same from competitors such as Google, there simply appears to be a greater tendency to stick with these once they have them.
As an example, at release it was quite apparent that Gemini 3 Pro, 3 Flash and Opus 4.5 had a similar, possible even the same frontend-focused training base bought in from a third party. But while Anthropic has had both smaller and larger improvements across each subsequent Opus release, I have not noticed such a consistent improvement with Gemini models over the last year. Maybe Anthropic simply buys updates more often, maybe they create more artificial PRs on top, who knows.
Same with the LIDAR scans, there was a point a few months ago when models started to exhibit a very systematic way of handling visual references in some few leather-crafting tasks, enabling whole new ways of working more efficiently combining existing measurements with images. Step-by-step pixel measurements, compensation for angles and poor captures, etc. Have since seen this in other models, again very much feels like a bought in data set that they are pushing as far as it goes.
Then there is 3D capability with models like Muse Wonder Full. I have not used it myself, if it ain't on API, it doesn't exist, but have seen PRs and do find the model output very parsable, despite it being terrible at any modelling task because these capabilities do not emerge from just training on vast amounts of data, they require high-quality, well tailored, step-by-step reference data of the entire process.
Harness wise, while pure output performance is very consistent across harnesses in my evals (I nowadays use both OpenCode via API and the labs own harness three times each to normalise), Anthropic also appears to be far more willing to use the time between releases to prepare features for the next. Stuff like workflows, proper subagent implementation, advisor, etc. often have found there way into Claude Code either in-between models or were first seen with a model that couldn't fully leverage these. Also feel that explained the good overall launch experience Fable 5 delivered, they simply used the time between announcing Glasswing and launching full access to get the UX/DX polished.
Meanwhile, I am still testing GPT-6.2 Astra and Sol so will refrain from making any more judgement on either till I have all data, but apparently there has been an extensive last minute rewriting of the system prompt to improve performance and "behaviour". Given the gap between GPT-6 Astra and yesterday, along with statements by OpenAI staff alluding that they had this model ready for a while, such things should happen in a far more structured, systematic, well thought out manner, not last second after some negative outside feedback. Heck, from what I know Anthropic, unlike any other lab, does zero early access outside the limited hardening for very exclusive partners like Glasswing. Certainly no early access for regular users, reviewers or the like from what I've heard, so Anthropic does all that prep-work for model releases with less feedback than other labs.
> Is there any theory on what is Anthropic's secret sauce?
I'm planning to write a blog post on it, but my theory is that it's aligned to not be "politically correct" but instead to have a solid stance. In the Claude Constitution, they explicitly say how they train it to have judgement and not "follow rules"
I suspect that Anthropic has more user data to train on from Claude Code. Claude Code became popular sooner than Codex and Antigravity, and remains more popular - thus more data to collect.
Cursor was popular before any of those three, and just look at how much Grok improved after they acquired Cursor's data.
OpenAI has plenty of feedback from Codex. Maybe not quite on the same volume as Claude Codex, but still massive. So I'm not sure that's the whole story.
I doubt there's a single reason for why Anthropic are ahead right now. Building a frontier model requires executing well in a wide variety of things. It's probably a result of being slightly ahead in various aspects of of data collection, model architecture, pre and post training, harness development, etc.
It's a shame that so much of what's happening is being closed doors. Maybe one day, decades from now, somebody (or some AI) will write a definitive history of this period, explaining what was going on at the various labs, how their research compared, and what the major players really thought.
They just have an incredible data flywheel I think. Since they were first to market with claude code, they have more training data than others. They also pay a lot ($150 per hour) for SWE experts to train the AI. I get some random recruiting emails occasionally from companies affiliated with them.
Yeah I’ve gotten the same thing. If anyone from Anthropic is reading this: it’s not enough to make me quit my job for a short term contract. Try doubling it and we’ll talk
According to benchmarks; time will tell. I see a lot of comments here and on X about how Google's caught up. But when Opus 5.5 came out, I didn't see comments. I saw examples of what people built, and it was mind blowing. Stopped the world for a few days for me and others, while we re-configured our brains around what was possible. I'm entirely willing to believe Google's catching/caught up but - lets have some real people build and show off like they did with Opus and Astra.
(After a long conversation with existing messages)
User Prompt: Ok, so SSH to the unraid server and update the plex db
Reply:
I need to think how to answer this.
The user is asking me to SSH into the unraid server. This seems to imply there is a machine to which we can SSH to. I should look for places where hosts may exist.
The implication is that once SSH is established to unraid, the plex db can be updated. It seems plex is a service that the machine has in which we can run updates.
I'm wondering if plex might be the Plex service the user has been discussing.
So the user doesn't want to SSH to the server - actually they want to SSH to the server.
... Formulating a plan ...
[some messages later]
Aha! THIS IS THE SMOKING GUN! `~/.ssh/config` HAS THE REQUESTED HOST!!! IT'S TIME TO CONNECT!!
[some messages later]
Ok, time to act. I'll ssh to the server and follow coherent rules.
----
I am shocked at how primitive its reasoning sounds (it reads just like gemini 2.5-pro's old reasoning). It often gets the job done (although it's quite slow and can't tackle more sophisticated problems).
Perhaps Anthropic's internal reasoning is also like this, but they really rewrite it to make it seem more coherent (which certainly sells it on me)? I almost want to click stop just by reading this nonsense.
I want to believe you, and it's what makes the most sense! But if that's the case, why is google (intentionally?) presenting such terrible reasoning traces? To make sure that if they're trained on, they're trained on garbage? I guess it makes sense, but still...
I was not really fabricating the example I gave too much. Here's a verbatim transcript from a session which was filled with this stuff:
... thought LET'S READ RecordReader.cs RIGHT NOW!! RecordReader.cs IS WHERE EVERY SINGLE TABLE, ReadPlan, DecideValueVisibility, ReadRowCounts , AND ShowValues READS ITS DATA!! Let's read RecordReader.cs (lines 1-300)!
It seems rather ironic that Anthropic givens the most logical reasoning traces, though. You'd think if they're ahead they'd either hide them or make them produce garbage (like what seems to come out of Google's models).
Neither has any relation to the actual reasoning trace. The summary is actively misleading and you'd be surprised how different it is from the raw CoT of the model. Just ignore it.
If you want to see example traces without jailbreaking claude or gpt, see open models that display the CoT without trying to hide it. It's pretty useful sometimes, with usual caveats (it deceptively looks like English but is heavily reward hacked, and it's generally a weak proxy for model's reasoning process).
This makes sense. Nonetheless, I find that reading Anthropic's reasoning traces is sometimes a trigger for researching more stuff. Often I'll see models thinking (or fake thinking, whatever) about something and I'll be curious to check out more about that topic if I'm not particularly familiar with it.
Off-topic, but this kind of conversation tells me just how little the average HNer still uses LLMs outside of specific coding tasks, or at least pays attention to them. Which in turn explains so many comments. Probably a good thing though.
You're of course completely right. Interestingly the "summaries" have only become more diverged over time. I now believe the big 3 are barely summarizing at all any more. I deem it likely that the small models don't get the CoT at all, they just completely make up something "plausible" solely based on the user message with a little bit of previous context. Much of the time it will be the complete opposite of something I asked for, or barely related.
Anthropic has changed this up multiple times over the last 12m, I'm talking about the iteration that was around the longest this year. I think they changed it again within the last feww weeks and potentially started actually summarizing again.
Your suggestion is the right one. Everyone who uses LLMs for their job should spend some time with open models like Deepseek and see what the thinking actually looks like.
Not only that, Anthropic injects fakes reasoning and tool calls in there to make it harder for distillers. It's shitty behavior but you'd do the same if you had an edge you wanted to maintain.
One of my long horizon tasks benchmarks involves writing logic controlling a team of soldiers in a simulated environment. Argon and Carbon both made such discoveries such as "it would seem that i need to be pointing at an enemy to shoot them". Its almost entertaining to read, its like its discovering the world on first principles.
(And it did produce a very competent algorithm, though it took a few days and a metric ton of tokens; I just triggered the same task with Neon).
Claude annoys me with condescension sometimes... makes it feel far less like a 'tool' and can even come across as a bit judgemental!
Just today I was filling out a security questionnaire and it said "I won’t write a PDF that claims zero credentials while this token is in the logs." But we were in the process of removing this from the logs so it was perfectly accurate to write and I wish it had deferred to my explicit request...
I think Anthropic overcorrected on the sycophancy; Google models to me seem a nice balance of highlighting issues / risks while still doing what I actually ask.
Anthropic models won't lie. When we fix security issues identified with sth. like Wiz and are writing a report, it won't write a summary suggesting that something we just fixed wasn't broken (even if that thing never made it to production). It has something akin to moral standards :-)
Claude models will go to extraordinary lengths to ensure sensitive data does not end up in their context window. It's a bit nuts. But frankly I appreciate it a lot. Even when it is annoying I know the model is doing the right thing.
Opus 4.x really struggled with secrets rotation in sops, even in 5.5 it's constantly emitting secrets to the shell history. Thankfully everything gets tested in dev, but Claude models are the reason why we have such good secret rotation now.
Yeah, I notice this as well. I was building something to help me understand logs from the different coding agents and it wrote redaction code to redact secrets out of the logs just in case if the logs had any so it wouldn't see them.
How do you evaluate opus 5.5 as "better"? My experience has been that Argon was ok - maybe equivalent to opus - but carbon in orchestration mode was finally able to rewrite some early experimental (ai slop) front end code - and break a bit out of UX/UI rut that a couple of projects were stuck in.
It might be that opus is better in better structured projects, and carbon is better at (re)structuring poorly structured projects?
Carbon is a better project manager/product owner - opus a better programmer?
> Is there any theory on what is Anthropic's secret sauce?
Lots of PR it seems?
I have all of Ant/OAI/OpencodeGo subs and test lots of models and I just cannot understand how and why some people find specific models so ahead, when it seems to me like reality is so nuanced and so specific to the particular task and how you approach it.
More like, most people especially on here do 90% of their "tests" on very similar tasks - coding on specific codebases - and the other 10% is just asking stuff in chat. So they never get to experience the jagged edge. Anthropic models could be horrific at translation to Nahuatl or greenhouse cherry tomato management compared to Gemini and OpenAI but there's one person on HN doing such things and tens of thousands doing enterprise SaaS coding.
Something that is interesting is how it's positioned in the overall fleet of Gemini models. Although overall more Intelligent than Carbon, Carbon is still positioned as the 'long horizon' model. This is clear on AA Index, where we see higher overall scores for Neon but lower scores in evals like CritPt which test research-scale reasoning capabilities.
With the frontier getting more jagged, we might see more of this 'less intelligent overall but useful for the frontier tasks'-type models, meaning picking a model will only get more difficult from here.
I do apologize in advance for putting you in spot like this, but there are two things I have been quite curious about.
1. Since 60% of the benchmarks contributing to AA are now private only, how do you guys ensure transparency? More precisely, how can someone verify the transparency in scoring?
2. There are also disclosures about custom benchmarking services offered to AI labs, what does AA do to keep the custom benchmarking services not getting access to the same/similar data that is used in the private benchmarks?
Okay, this "better at long horizon tasks" thing might explain why Carbon scores worse on most benchmarks yet is the larger model.
I looked into this using data I have from a battle system built on top of OpenCode (so we're trying to capture real world use). It looks like Carbon actually does a lot better when it uses significantly more output than opponents.
Win rate when output is less than 2.5x opponents: 53% (41/78)
Win rate when output is more than 2.5x opponents: 73% (41/56)
There doesn't seem to be any relationship between actual size of task though. Win rate appears random based on total output of Carbon and also opponents. (Though our data is still small, so it might exist.)
Looks basically like Carbon was trained to think a lot longer, not necessarily that it's better on large tasks.
Yes, I feel the next differentiation in model capabilities would be wrt how 'long of a horizon' goal they could autonomously work toward without breaking. We're in need for a new generation of evaluations and benchmarks for this -- METR doesn't cut it anymore.
This comment always comes up but isn’t this literally irreducibly complex in the sense of Kolmogorov complexity?
I’m out of my depth here but it does seem prompts have a “smell” to their complexity that means we can use a classifier to mate them to a model with better success than random noise, but still don’t essentially all of the savings associated with systems like this just come because there is some “complexity arbitrage” where there is enough of a token cost differential between models that you always save money by trying the cheap model first and then trying a hard model if it turns out to be hard?
There's stuff that does this and I've put a substantial amount of time last couple months looking into it, unfortunately you always, always, end up at chicken-and-egg problems.
It's really frustrating to describe. Easiest is you can't score on just the prompt, you need the full context. Ex. "Solve Fermat's last theorem" initial prompt then "continue" followup prompt. Which gets you back in the tarpit of "if I have to pay a model to do an inference run on context, then, the context is cached and I should continue using that model."
The vision of another model thats RL maxx'd on knowing how to delegate has a few problems. One, we can observe externally, it died/didn't work even at OpenAI's scale and willingness to make it work, i.e. the...GPT-5.0? IIRC...router that silently went away. The other, if such a model was built, it'd also be good at regulating it's own thinking enough that it'd naturally use less tokens.
In practice, the frontier hackers have settled on a decision model (i.e. Jev) x calibrating effort level or choosing from a set of models x reasoning effort tuples they specify rules for choosing in their Jev prompt. My guess is this stays stable and the only real choice.
I suspect they may want to get away from the size class due to the fact that - at the pace they are launching new models - and the similarity of model sizes between Neon/Carbon, Opus/Fable, Sol/Astra - the best model is going to be the most recently released one, not the biggest one.
I'm not sure it's much better than 4.1 Flash vs 4 Pro -> if 4.1 Flash is better temporarily when Pro is typically significantly better, people are going to be confused.
I'm not really sure how you get around this problem easily.
It's not going to age well when the next Flash-lite model launches, but maybe they will call that `Gemini 4 <Element>-lite` - who knows...
Interesting. You're claiming that the Opus/Sol/Neon class and Fable/Astra/Carbon class models are similar sizes to each other? That would be surprising to me.
I've found that while Opus 5.5 is often pretty nice for practical work, Fable (even 5.0) is still significantly smarter on hard tasks and makes far fewer "dumb mistakes"
A Microsoft Azure page briefly claimed Sol runs with less looped transformer passes; and so you could speculate that Sol and Astra might share weights; but have different “loop” factors.
That would make a lot of sense from an inference optimisation perspective. The same prefill and decode machines can hold one set of weights, and serve both Astra and Sol requests.
At (multi-)terabyte-scale model weights, and when you have to offer not just the latest like Fable 5.2 but still serve 5.1, 5.0, etc; heck I think Opus 4.6 is still around; you really care about overall compute utilisation.
Idle GPUs are wasted GPUs; and even over the fastest interlinks you’re not hotswapping multi terabytes of weights.
With MoEs, there’s probably stuff you can play around with varying the number of active experts; or other clever optimisations.
Maybe Sonnet 5.5 could be just Opus 5.5 with a different router and less active params? Who knows; I’m just theorycrafting now :)
> A Microsoft Azure page briefly claimed Sol runs with less looped transformer passes;
This doesn't really make sense given how the models released.
OpenAI released Astra long before latent-space looping was a thing at the frontier space.
Then they released Sol 6, which was fast, clearly not looped. Then they release Sol 6.1 which was nearly as good as Astra, but way cheaper, and WAY slower -> this would follow it just being a looped version of Sol 6.
It doesn't seem likely that Astra 6.2 is just a looped Astra 6... It's too fast, and they'd go bankrupt if they're charging the same price... It may be a further looped version of Sol than Sol 6.2 - which would explain why it's not as much better as we would've hoped.
It may also explain why their release is not going as well as they hoped. It's hard to predict how your first looped models are going to fare.
If you look at how long it took Google to recover from their failed 3.5 Pro, it would be surprising if GPT recovered from their rumored failed run on Astra in under a month. That would give them reason to release a looped Sol and call it Astra.
I think they're responding to the claim Astra can't be looped, there's clear and convincing research, as well as public comms, that it is. It's unclear why Astra can't be looped because looping is new, but, 6.1 Sol released a week later can be.
Astra was the first shipping looper. Unconvincing evidence would be reporting x OpenAI's quick reaction to it via a blog post from their chief research (it broke what's considered a safety taboo). Convincing evidence would be research from ex. Neel Nanda on serial reasoning, recent update at https://x.com/EpochAIResearch/status/2122762411287748696?a=J...
I heard rumours that Sol 6.1 was cut down Astra because Sol 6.0 was received badly. But I thought that was the exception (and a result of Open AI's actual Sol-class model being uncompetitive) rather than the rule.
That would line up - on release and for the few following weeks Sol 6.1 felt like Astra but maybe more task oriented, recently it’s been a noticeable downgrade in the level of specification needed to successfully set off.
Fable & Opus, Astra & Sol, and Carbon are all thought to be >1T param models - AFAIK, most people think Opus & Sol are ~5T param models and Astra, Fable are 10T+.
I think the rumor-mill consensus is that Google's models are significantly smaller while attempting to swing at the same level - but the ratios AFAIK are thought to be around the same - i.e. Neon is somewhere around half as big as Carbon, big enough that a newer checkpoint can make it temporarily a little bit better.
There's only a limited amount of hardware for training and serving models at scale, and which form an upper bound for what model size/inference can support.
A GB300 rack has 72x288GB=20TB VRAM. Do the math and 10T parameter models are just about the max size you can easily train and serve (at 1M context) without doing a lot of custom hardware development or limiting your scale.
Honestly, the bigger question is, are they at some point start adding numbers (Carbon 2, Neon 1.5, Hydrogen+ ionized), or are they keep giving a new name to every single new model. Argon seems to have been basically "replaced" by Carbon.
Why does this chart not include the other gemini 4 model, Argon? Also I am sure somebody thought about which element to name which model and it would be very helpful to me if they would explain their reasoning. OpenAI's naming got kind of abstract with 5.6 but I could follow it: Luna(Moon) was smallest, Terra (Earth) was medium, Sol (Sun) was biggest. Astra kind of breaks from that but whatever. But like I would genuinely appreciation someone from google to explain the reasoning behind their element names. Like are Argon and Neon related in some way because they are both noble gases? Is hydrogen the smallest model because it is the lightest element? Carbon has a lower atomic number than neon but is more expensive and presumably bigger... maybe I am overthinking this.
Inconsistently it seems. On the $20/mo tier, Argon has been removed from the web interface and app, but is still available in Antigravity (while Carbon is only on the $100/mo tier in both).
I have to assume they're messing with something other than size here - nobody would be this bad at naming different model sizes. Maybe sequence length or post-training strategy or sparseness or something else with the architecture.
(Although why they'd release all ~four is a mystery. I guess they're having a hard time picking the "best" one and so are just putting them all out to let the public decide.)
“Model size class” is a completely meaningless taxonomy.
It doesn’t correlate with intelligence, speed or price furthermore you have no way of telling what the model sizes of “like for like” named models from oai or Anthropic are.
But Argon has a higher atomic weight than both, and sits below Carbon and above Hydrogen. Also, Carbon is rated at better at long-horizon tasks while Neon seems to be smarter at contained work. So they're not strictly better/worse.
That's my thought too. If one needs a lookup table to understand the product by name then it's a product issue, not a consumer issue, even if everyone "should" have the first N elements memorized by atomic $property.
Ugh, this sound so overly nerdy, that it could be very true. I do not follow or care about the Google LLM releases just because the naming scheme is so convoluted (and the constant churn of products around the llms). I learned the anthropic naming scheme, it is stable and easy to follow. And this is why Claude is my daily driver. I don't like change, I like stable and predictable progress.
Open AI naming scheme is kind of easy to learn (bigger planet = bigger model). But fuck Googles bullshit naming crap, seriously.
I've no clue whatsoever how to read anything about the names of them and what they might imply you would use them for though. Within these most of them have LOW MEDIUM HIGH XHIGH, with the default being different for some of them, how anybody is possibly meant to understand this whole matrix of ambiguously named products is beyond me.
Same here. First i thought it would be noble gasses. Then I thought it would be about the atomic weights. I know, I should just ask AI to figure it out for me, but its a bit humiliating.
Maybe this implies that Neon is smaller than Argon in weight size, but it's newer, but it's also only a little bigger than Carbon? Do the Nobel gas models have more safeguards? Is Carbon 6x the size of Hydrogen and a third the size of Argon? When they release Xeon can we make compounds? Feels like some really cheap Da Vinci Code esque thing if we're supposed to get some sort of meaning out of the names.
> In our early evaluations, it matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost. We’re excited to bring these capabilities to Glean customers tackling multi-step work that demands accuracy, detail, and speed.
So possibly smaller than Carbon? Or some other axis that is making it faster and cheaper? No idea.
As far as I can tell it's arbitrary and Google is just choosing element names they think sound good.
When Argon first came out it seemed like they could be going for noble gasses, like Helium (small), Neon (mid), Argon (big).
But then Carbon came out and it seemed like maybe they were going through the alphabet for model generations. Argon, Boron (internal), Carbon.
And now they've come out with Hydrogen and Neon, and there's no logical way to order them. The actual model sizes are Hydrogen (small), Neon (mid), Carbon (big). This doesn't match sorting by atomic number or alphabetically.
They are almost certainly being ambiguous so that people rely on the model descriptors to pick which one to use.
I think there are a ton of people who automatically default to the largest model with the highest thinking for everything. Which historically made sense but not nearly as much anymore.
At this point a model like hydrogen can cover just about everything an average person is doing. You don't need Carbon deploying an agent swarm pushing terabytes of parameters through HBM memory to help you plan your birthday party.
They use the very precise "it sounds cool". Neon is 75% the price of Carbon but has 4 extra protons in the nucleus and is also a noble gas. Argon is a heavier noble gas at 16 protons.
element/model - atomic number - cost
H - 1 - $1.50 & 7.50 per mtok
C - 6 - $4 & 20 per mtok
Ne - 10 - $3 & 15 per mtok
Ar - 16 - $2 & 10 per mtok
Claude - length of story but fable messes it up
Haiku - 3 lines
Sonnet - 14 lines
Opus - a novel?
Fable - 30 lines??
Mythos 15700 lines?
Luna - 28 days
Sol - 365 days
Astra - 5-6000 stars
Everyone seems quite upset/perplexed by the Gemini naming scheme but I quite like it. And hey, if you find yourself lost, the periodic table is probably the most ubiquitous ordered list in the world and serves to clear things up pretty fast. This will likely not be strictly true as models release over time, but is a good rough indication of what the team thinks we should expect of the models, in the same way the other labs run their naming.
I also like the personalities implied by these, with Anthropic saying “human literary achievement is neat”, OpenAI saying “space is neat” and Google being like “science is neat”.
Edit: I was coming back to say “I wish we could talk about stuff more substantive than the naming scheme” but (1) my comment may be objectively incorrect, and (2) I have to apologize that I haven’t helped the situation by… commenting on the naming scheme.
They're in alphabetical order, almost like Ubuntu release names. There was a leak about a Gemini Oxygen model, which would further support alphabetical order -- skipping names they don't want to use.
So, names don't align with capability but instead "release freshness". This is unlike every other provider but it is what Google does (did?) with Android releases I believe (dessert names).
> And hey, if you find yourself lost, the periodic table is probably the most ubiquitous ordered list in the world and serves to clear things up pretty fast.
But the way google ranks the models by capability doesn't follow periodic table order.
Hydrogen - 1
Carbon - 6
Neon - 10
Argon - 18
It could very well be based on their density (g/cm^3 in standard conditions):
Hydrogen - 0.0000899
Neon - 0.000900
Argon - 0.001784
Carbon - 2.267
This would make more sense than assuming it is based on their periodic table order as it more accurately maps to how Google markets their capabilities, but is too esoteric to really be useful for clearing up confusion for normal users, IMO.
The periodic table isn't really helping. Hydrogen is the lightest, that much makes sense. Beyond that it doesn't seem to work. The elemental ordering is H C Ne Ar. The cost ordering is H Ne Ar C (with Ar and C being the same per token but I gather that C thinks more and thus costs more). The capability ordering seems to be H Ar C Ne. Argon is the highest on the table but is neither the most expensive nor the smartest.
It looks like it's being served extremely fast via Openrouter right now https://openrouter.ai/google/gemini-4-neon. 127 tokens per second. This is way faster than Carbon at 46 tokens per second (https://openrouter.ai/google/gemini-4-carbon). I'm a bit confused by this, since it seems like Neon is supposed to be their flagship model. Perhaps they've been throttling Carbon?
Carbon has a huge problem that it's just way too slow, and outputs way too much per task (on real world tasks, it's the worst at this https://micjm.com/blog/carbon-is-that-slow/).
Without the speed issue, Carbon would be a top model already. It's top on our battle system built on opencode, if you exclude early votes: https://agentduel.com/leaderboard?excludeEarlyVotes=true, though to be fair, small sample size so far.
Conclusion is that it's looking good for Neon. Even if it suffers from similar problems as Carbon, if it's even smarter and way faster, it should be a top model, maybe the best.
Edit: I was wrong about Neon being the larger model. I made this assumption because Neon has higher scores and they're both "v4".
Okay, that clears it up. Their naming convention is confusing because Neon has a larger atomic number than Carbon, so there's just no pattern here at all.
Weird because Neon is outscoring Carbon in AA index, Terminal bench science, etc and they're both named "4". Should have released this as Neon 4.1.
With that clarification, I'm sure Neon has the same overthinking problem.
Google (and the other labs) don't give raw model sizes. It's inferred from price/speed/benchmark/quota. When you align those parameters, it's clear what models are competing with what.
Also neon seems to be stepping on the toes of Carbon a bit, just like Opus 5.5 did to Fable.
I'm almost positive google is purposely trying to breakaway from names describing sizes. If it works (read: helps balance demand), then the other labs will probably do it too.
it would be funny if they did some marketing around the hydrogen model (like as powerful as the h bomb) to try to push more non coders to use it for normal tasks
I've been incredibly impressed with this model so far. Argon felt like a leap in presentation, the model writes really good documents and replies. Neon feels like another big leap in usability over that, the model seems pretty smart but its super eager and runs down tasks really well and efficiently. Definitely passes the vibe test
I am a Claude Ultra subscriber that's very happy with Opus 5.5.
If I wanted to realistically try out Neon in some of my workflows without committing to a paid subscription before seeing if it can actually work as well as Opus, do I have any option, or do I just need to take the gamble?
I think that page is out of date. You can use it via T3 Code now and Google has said it's okay. It's via their official ACP-compatible endpoint that's now provided by the antigravity-cli.
You don’t need to use the models through antigravity. You can get them through ai studio (or whatever Google calls it) and pay api pricing to use it via whatever harness you want.
They say it's going to be available for Pro users - so you don't need a $200 account (Ultra). But, I don't think you're going to be able to try it unless you get at least a $20 account.
IIUC, you also don't get to try Opus unless you have a subscription, right?
Reminder that Google trains on your data if you're using a consumer sub whereas Anthropic and OpenAI claim they don't if you've got all relevant boxes turned off. I know everyone stopped caring about this at the start of the year and believes the latter two don't honor it anyway, but it's good to know.
>how can one presume the sun will come up tomorrow merely because it has come up every day prior? have some epistemic humility! i am a very smart rationalist.
Edit: I tested the model and it found several mistakes Astra 6.2 made. It’s possible…
This is an interesting model - I feel like they're trying to avoid OpenAI/Anthropic's approach of various model tiers, and just keep climbing the Pareto graph (for lack of a better term) as fast as possible, both for cost and performance. I think I personally like the approach, despite the confusion it causes.
Major props to Google for announcing a model and releasing it on the same day.
Unfortunately, their recent change to allow me to opt out of AI training (which would also include connector data like Gmail, photos, etc) hasn’t rolled out to me yet[1]; so I won’t be upgrading my subscription; nor using the Gemini app until I can opt out.
> Currently, if you disable history, Google still retains that data for 72 hours (and 24 hours for the user). Once the updated page is live, this duration will change to two months.
Oh know, they're going to steal the code for you 3 MAU SaaS !!!
/s not trying to be mean, its just weird hill to die on. Every company is stealing all your data in ways you unknowingly consent to at every second of the day.
Anthropic competing with Google instead of OpenAI at the frontier is unexpected, I have been doubtful of Google DeepMind's status as a frontier lab in the past few months, but damn they've been cooking recently
It's important to note that DeepMind and the other two labs aren't playing the same game.
The latter need to raise funds and IPO, but the former is an arm of one of the most powerful companies to ever exist with some of the largest data warehouses around and custom hardware. And they have an endless pile of internal code to maintain already with hundreds of millions of end users already bought-in.
The incentives being different leads to different focuses and product lifecycles.
Just ran my data analytics benchmark [1] on the model: Excellent results for this model like the rest in the Gemini 4 series (all models score an A).
Costs spread between Hydrogen ($6) < Carbon ($11) < Argon ($13) < Neon ($17) for this 43 question exam.
At a quick glance, the tone is good as well. But we do a fair bit of prompting in our harness to guide the model, so I can't say how the default model behaves.
Carbon 4 has been incredible for my use, tuning a local inference engine based on tinygrad running on Tenstorrent hardware. I had a barely functional prototype I built with Claude that petered out at an unusable level of performance. Astra and Sol didn't make much progress when I pointed them at it. Carbon 4 more than 10x'd the prefill and doubled the decode performance on Qwen 3.8 27B without breaking quality or the generality of the stack. Very nice to see competition at the frontier again.
I'm running 2x p150a's right now. I have them hooked up to a mini-ITX PC with a single PCIe slot via a quad x1 switch (figuring out power to wire up the other two cards I own). My best numbers so far are around 2500t/s prefill about 52t/s decode with DFlash2. All of this is vibecoded besides the original design sketch which was maybe a 3 line description of what tweaks to make to tinygrad and which parts of the existing TT stack to use (almost nothing).
If one day, AI becomes provably good at all programming tasks, at working at mathematics better than all mathematicians, and at solving problems in almost all domains with maximal efficiency, I think it'd be a dystopia.
What I simply don't get is why people want to push the boundaries of efficiency and knowledge. I used to think as a teenager that all knowledge was good but how good is knowledge when we can't acquire wisdom with a comparable speed?
The greatest thing about the early internet and earlier science was that people shared and communicated with each other. Now they just share what they produce wihouth much humanity at all.
It seems to me that the creation of all these models has just become a race where no one really cares about the finish line and no one really understands why we're doing it, except for the short-term trail of dollar bills on the race course.
For me, each new model release is a sad day for humanity.
Does anybody have a corpus of examples I could look at? I've wondered if it's worth setting up Gemini as a human-facing documentation author with other models still being the underlying workhorse.
It's a very good model. If they get their UX right, it has capabilities that Opus 5.5 and Fable could never have (e.g. watching YouTube videos, generating images).
I have the "Google AI Pro 5TB" plan for Students. I see "4 Hydrogen" and "4 Neon" in the model picker on gemini.google.com. When I look at one.google.com's plan details, I see "Your Plan: Google AI Pro (5 TB)". When I click "View Benefits" on that page, one option in the list says:
Google Antigravity
Entry rate limits to advanced models, including Gemini 4 Argon and Opus 5.5, in Google Antigravity, our agentic development platform
Yet, here I am, running the `agy` cli, and the model picker says "Unavailable: Gemini 4 Carbon/Argon/Hydrogen/Neon is currently available to paid Pro and Ultra subscribers". It has been this way since the release of Carbon. I guess it's "3.8 Flash" or nothing for me.
I had to grab my laptop to say, you are actually incorrect. I have been using Argon for the last week or two via Antigravity on the Google Pro $20/mo plan, discounted to $5/mo for 3 months, just fine. And as of today yes, I see Neon and Hydrogen there. The only one I can't see/use is Carbon. https://i.imgur.com/2c9EnbV.png. This isn't just a visual thing, I have actually used every model there. They all work.
I think that its likely the reason for
a) the strange naming scheme
b) the promptness between releases
is that google has outpaced internal expectations for model development ( probably due to RSI ) and now can now utilize its vast compute pool much more effectively for autoresearch and exploratory self-improvement tasks. I wonder if this suggests a steeper trajectory for them than OAI/Anthropic in general.
Well, I use my product as a benchmark for this new models and I have to say that Argon is still the most efficient model by far for my use case. Carbon, Hydrogen, and now Neon all use 3 to 4 times higher tokens compared to Argon for creating demos. Neon was the first Google model that passed 8192 MAX OUTPUT TOKEN limit, which is disappointing a little bit. But lot's of experience I got from these models, for example bigger models like Argon spend significantly less output token so even though unit cost is higher both total cost and the output time can be lower than the lower models.
> Independent evaluations rank Neon as one of the top models and on benchmarks like the Vals Index, it ranks #2 (70.9%), and reaches #1 on both the Legal Research Bench and Tax Agent Bench.
On Artificial Analysis Intelligence, it scores comparably to GPT-6.2 Sol but is significantly more expensive.
Okay so it seems like the ranking here is Hydrogen -> Neon -> Carbon even though that is not clear at all from this release. At least Googles model picker in the Gemini app now lists those 3 above each other in this order, and keeps Carbon for “Code & complex work”
So Neon is the new default, and Carbon the slow deep thinking “when you need it”? A bit like how Opus 5.5 was better than Fable 5.1 in most things, but Fable still had the edge over deeper things (then 5.2 came along)
Hydrogen has already became my default for 80% of small tasks. If Neon is as fast but smarter that would honestly be very cool
To be fair prior to Gemini 4, the app had 3.1 as the "Complex work" model while 3.8 Flash, far superior in every way to 3.1, was listed as the "Fast" model.
So I'd take their hierarchy with a grain of Sodium.
This is good, it makes me have to understand the full names of every element in the periodic table.\
(PS. Taiwanese)
But I still can't understand the differences between these models.
Recently I watched Veritasium talking about helium, and I was thinking, will the effort setting of these models eventually have something like "Temperature K" to configure?
I think for large refactoring jobs this will open up a new category of work that an agent will be able to do. Also reasoning steps count towards output tokens, no?
Looks to me like just another math benchmark similar to FrontierMath.
> We introduce Riemann-Bench, a private benchmark of expert-curated problems designed to evaluate AI systems on research-level mathematics that goes far beyond the olympiad frontier. Problems are authored by Ivy League mathematics professors, graduate students, and PhD-holding IMO medalists, and routinely took their authors weeks to solve independently. Each problem undergoes double-blind verification by two independent domain experts who must solve the problem from scratch, and yields a unique, closed-form solution assessed by programmatic verifiers.
This seems to be specific to Gemini because I have not experienced this with OpenAI's models. In fact, Codex now has a feature such that the model can ask a question and continue making progress on work that is not blocked while waiting on your answer.
I wasn't benchmarking or anything rigorous, but I noticed it with GLM 5.3 Flash. It was downloading and installing stuff to my computer to try to work around an assumption in the prompt that it knew was wrong. I also noticed it trying to work around websites' bot prevention measures. I stopped it often.
They have just removed 3.1 Pro and Argon (and Flash) from my Pro subscription without any warning. I was using 3.1 Pro for some important workflows. It was actually better than the newer models for some tasks. Now there is suddenly only "Auto", "Hydrogen", and "Neon"?
Maybe the people at Google don't realize that but some people need constancy. If a model was good enough for some work 6 months ago, it's not suddenly becoming not good enough for the same task.
I'm pretty sure Google is intentionally using non-descriptive names to get people to use the model they think is best for their task, because at this point, it probably is.
On Google's scale you could probably mint a datacenter of free compute if you could get people to stop using 10T parameter giants to vibe check their emails.
As someone who works with real neon signs, the "Gemini 4 Neon" image at the top of the post is so unrealistic and jarring. It reads to me like a seven-fingered hand. Probably won't bother most people but it was a funny first impression.
Okay, but now will you let my tiny team sign up somewhere to be able to use this the way I use `claude` or `codex`? I'm told to use `antigravity`, but when I go to https://one.google.com/ai (which is where the magic answer from Google search told me to go), I'm told that I have to use my "personal" account, not my work account. So I clicked "Manage Google Workspace Account" on that page, and got dumped in a super dense UI with nothing related to enabling this model somewhere.
So I haven't tried this, but it looks like you have to set up the Gemini API in a Google Cloud project, then point your Antigravity at the project. https://antigravity.google/docs/pay-as-you-go/
Thanks for finding that. Looks like it would use API pricing? I was hoping to be able to get a subscription like offered by the other competing models, but might not be able to do that.
The thing that confuses me with Google models is I have a Google pro subscription (even the names confuse me, I have hydrogen and neon in my Gemini now!). But do they even have a coding harness, is the coding harness included in the plan? Their naming marketing and generally everything else about them is as confusing as possible.
I do think they're very well positioned, their small models are smart and fast, and integrating well with Google is a big plus for basic use case stuff. I always like these models!
So Anthropic is being valued at around $1.2T and their only revenue stream is a collection of frontier AI models with 0 moat that are basically as good as Google's, a company which has the luxury of other revenue "streams" (rivers/seas/oceans) and can certainly win any price war with Anthropic.
Everyone uses something you can remember somehow than there is google. Not only do you hear their model names rarely (neon, argon etc. is new right?) but then they choose something which is not obvious at all.
2027 january RSI is achieved.
2027 february first big scary incident (100 million people severely affected by infrastucture collapse for several days)
2027 march a fleet of rogue self replicating agents is discovered.
2027 april first severe fatal incident (1000+ people dead, most likely from a missile created using oss models).
2027 june. Deadly virus created by a terrorist organization is released in Ukraine.
Remind me!
something tells me this janky naming scheme will mean the whole gemini section of google is just gonna get shut down and replaced with something new during 2027
The names just keep getting more random. And the table at the bottom, I'm pretty sure it's conventional (at least in the English speaking part of the world which I'm not originally from myself) to put your plan/model/price/whatever tiers left to right in increasing rate order. They do right to left here. Are they deliberately trying to confuse people? I get it that we tech workers are generally wary of google, microsoft, etc. naming conventions often (always) being terrible, but the average user out there (which increasingly is everyone with a smartphone or PC) could just skim through the announcement and go away with an easily avoidable misunderstanding.
My guess is that Google thinks the trade-offs are too nuanced to have a simple small/medium/large ordering. Neon is between Hydrogen and Carbon in terms of speed and cost, but it scores highest on Artificial Analysis Index and has a xhigh thinking level (unlike Carbon).
So now we've got to get a periodic table out to understand Gemini releases? Hydrogen is the weakest, and Argon is the middle, but Carbon is super-Argon and Neon is better than both? And I can't even look at the price list if my logged-in google account doesn't have AI Studio(??) access?
e: No, Neon is the cheaper, mid-tier between Hydrogen and Argon. So Argon/Neon are aligned along the noble gasses, and Hydrogen kinda makes sense as the weakest ... but they threw Carbon in without any consideration
And naming things remains one of the hardest problems in CS. At this point, I have no idea what to expect when I hear the names.
Just give them a version number and an identifier, like:
1. V5 smol
2. V5 mid
3. V5 big
4. V5 yuge
WTF are Hydrogen, Oxygen, and Nitrogen? At least OpenAI and Ant have Luna, Terra, Sol, Astra and Haiku, Sonnet, Opus, Fable. They use those consistently.
Luna->Terra->Sol->Astra is fine, you need to be familiar with meaning of the names but it makes sense and works.
Haiku->Sonnet->Opus->Fable/Mythos is better, more common words, more obvious understanding.
Hydrogen->Argon->Neon->Carbon (is there a Barium too somewhere? ) fck all I have no idea, people don’t even know noble gases well, but hydrogen and carbon doesn’t even fit that. 1,2,3,4 would be a better naming scheme.
I don’t know what’s going on with those GPT 6.2 prices. There should be some standardization auditable in the x-axis cost per task. If not, labs may game that through loss leadership tactics.
An army of Anthropic shills is going to whine "benchmaxxing".
That sounds like serious benchmaxxing for Gemini 4 Neon is ahead on 17 out of 19 benchmarks vs Opus 5.5.
Not so bad for a company that, according to HN, "has lost the AI race".
P.S: I like the Google models but a truce before the Anthropic IPO would have been nice. This is just a very big dick move a few hours ahead of Anthropic's IPO. Shall we see Anthropic launch a new model from now 'till the opening bell?
I think internally, Google, Anthropic and OpenAI are thinking "It's so easy, you just use Neon for X, Carbon for Y" - but most of the world is trying to map Sol, Astra, Neon, Opus, Fable, Mythos, Carbon, Haiku, etc.
I just hope they finally land on a naming scheme they are happy with and just stick with it. But I have my doubts. Perhaps they are all benefiting from this confusion.
Because probably there is no system. They just train a model, and find out that it's "kinda good", and that's it, they publish it. The models do not constitute even a partially ordered set, nobody can predict their ranking for your particular tasks.
Since I was curious about the origin of the name, I wandered over to where three Sony employees were chatting passing time until the day ended. I learned all three were mid-level marketing staff based in Sony's New Jersey US HQ. So I asked "How do you pronounce the computer's name?" All three had different takes between various long vowel vs short vowel sounds.
Since they were all facing me, I also asked, "Without turning around, how do you spell the name?" They all paused, then laughed. They did manage to work that out after a few false starts but it wasn't exactly smooth. As we chatted some more I asked who'd come up with such a unique name and learned Sony had hired a "really cool" high-end branding agency who specializes in coming up with great product names. The effort and expense in creating the name and logo had clearly been marketed to employees as "cutting-edge cool." I warmly wished them the best with the new product before wandering out of the hall to dinner.
Regardless, the PC and it's neat industrial design plus Sony's brand and distribution were good enough to overcome a weak name. VAIO was an acronym for "Video Audio Input Output" (which Sony later changed to "Visual Audio Intelligence Organizer", although I'm not sure why they thought that would be better).
Haiku < Sonnet < Opus < Mythos/fable
I think openai has a similar system with astral bodies:
Luna < Terra < Sol < Astra
Although, maybe it's called Fable because it's trying to bluntly convey strong moral lessons, hence the strong safety classifiers...
You could argue that's a detail the user shouldn't have to concern themselves with, but I think it's appropriate to expect them to learn that and pick names which put that concept front and center. The various "model router" approaches for consumer tools like ChatGPT have been finicky and inappropriate for professional tools we spend a lot of money on.
"Number four, obviously!"
So they fucked up naming as well in this regard.
Fable is Mythos with safeguards, basically.
In fact, Mythos also includes safeguards, so at this point it's not really clear what's the point. You can apply for security program and there are three levels of access to the Mythos. I guess now it makes sense to just drop "Fable" name and go with Mythos which will just be implicitly restricted according to your access level.
Or maybe they'll realize that there are abliterated openweights models, and restricting Anthropic models does more harm than good. I literally have to use other service for security audit of my software, because Fable rejected it.
Then the Trump administration panic-banned Mythos and Anthropic smartly decided to maybe call it something else.
But the naming isn't completely random, here's what I use to make the ordered list in my head:
Claude is named after creative works, and uses bigger works for more powerful models. In order of strength:
* Haiku - short poem (we all know this one)
* Sonnet - "medium" length poem (many of us know this one)
* Opus - It's the singular form of "opera", so it's like a stand-alone segment of a larger body of work. (Nobody knew this one, but not hard to remember, and know you know that "opera" is actually plural! The description sounds bigger than a sonnet, but it's not a strict ordering.)
* Fable - a story (this one doesn't seem obviously bigger than an opus, but I guess an opus is one part and a fable is a front-to-end story, it works)
* Mythos - another type of story, which works and it's basically Fable but Anthropic's internal-only version
Gemini, as of version 4, started doing something with atoms:
* Hydrogen - smallest periodic element, the most common gas in outer space
* Neon - a bigger element, less common in outer space
* Argon - yet bigger element, yet less common in outer space (as best I can tell, I am not an astronomer)
* Carbon - this breaks that pattern, its a smaller element than the first three, and not a gas... but critical for life on this planet? And it sounds like a longer version of Argon. So I can remember it, but I'm not sure what the intended pattern was.
The model ordering is not strictly upgrades, but AFAIK they are intended to be steps up in terms of (claimed) complexity and reasoning power. But since it's an evolving field, the specifics of how cost, speed, and accuracy combine can make it hard to know how to order a given model for your use case. (i.e. cheap & fast with a harness that makes more calls, or expensive and slow but better at one-shot).
Deepseek's naming is more organized and literal, but you still have to remember the prefixes for V = version (plain) and R = Reasoning, and that they have different specialized models that aren't intended to be compared to each other.
* Neon and Argon: noble gases commonly used in gas-discharge lamps[1].
* Carbon: used in arc lamps[2] (the first practical form of artificial electric light).
* Hydrogen: the sun (abundant light, available to everyone).
The symbolism of light would make sense for AI. Shedding light on a subject, enlightenment, a light bulb moment, etc.
If this is right, then maybe we'll see krypton, xenon, mercury, sodium, or tungsten next.
---
[1] https://en.wikipedia.org/wiki/Gas-discharge_lamp
[2] https://en.wikipedia.org/wiki/Arc_lamp
OpenAI's naming scheme is in Latin:
* Luna - Moon (smallest)
* Terra - Earth (medium)
* Sol - Sun (large)
* Astra - Star (could be larger than the sun I guess)
Meta's Muse:
* Ember - (smallest)
* Spark - (medium)
* Wonder - (large)
A spark is the beginning of an idea whereas wonder is associated deep thought. But an ember and a spark are both small pieces of fire, so the ordering is a little muddy.
Google's naming scheme has the least clear pattern of any lab.
But trying to keep track of the model names in all 3 companies? Utter madness.
We need something better. All I really care about most of the time is whether I'm using the cheapo, average, or expensive model. I don't really care about which actual model is occupying which actual spot at which time with which cutesy name. I just know I have a budget, and so I allocate the three levels according to the type of work I'm doing.
Anthropic: "We'll name after the size of literature style! An opus is obviously bigger than a sonnet!"
Much of the world: "What is this word Opus?"
OpenAI: "We'll use the Latin names for astronomical bodies. Everyone knows that terra = earth and luna = the moon and that the earth is bigger than the moon."
Much of the world: "Does terra mean it is terrified? Oh well at least it is better than Gpt 4o vs Gpt o3"
Google: "We'll name it Bard! Oh, that was bad. Lets name it Gemini because that sounds cool. Oh we need different models, ok we'll have Pro for the big expensive on, and Flash for the fast one. But now we want a fast large one and also we have mid sized models. Ok, lets name them after elements, but not ones that anyone knows anything about so now we can confuse them as to if Neon or Carbon or Argon is better. But Gemini Hydrogen is the cheap fast one"
Everyone: "Ok, Gemini Hydrogen is cheap. But this is too confusing, I'll just use Claude".
0 - https://knowyourmeme.com/memes/poob-has-it-for-you
I do note that gemini.google.com has replaced Argon with Neon, so I guess maybe that's a clue?
My personal approach is to not really care at all. These specific models aren't going to be here in 6 months, so learning the intricates of these specific models has no benefit for me. I'm just going to mostly rely on whatever one is the default
Expert models: Lion, Shark, Whale
Flash models: Swift, Falcon, Robin
Though I still think it shoudl jsut be : Expert v13 , Fast v3
Yeah grande(big?) and tall are whatever to your point
While cutesy, most current models are quasi-semantic progressions.
Claude (length, or cultural durability): Haiku, Sonnet, Opus, Fable
GPT (altitude): Terra, Luna, Sol, Astra
---
Gemini unfortunately has no semantic key Hydrogen, Carbon, Argon (superseded by Carbon), Neon.
Camry = Kammuri = Japanese for "Crown"
Crown = English for "Crown"
Anthropic is the new Mercedes? OpenAI BMW?
This wouldn't be a conversation if the big 3 didn't change their model names multiple times per year. Or conversely - do you think people would be fine with Toyota renaming the Camry every 6-9 months?
I feel like Google purposely names their models not for humans but for AI agents. People keep saying: "make your products ready for agents," so maybe that's what Google is doing, which is making Gemini 4 models ready for other agents to use.
Android's silly food naming convention works when you do one release every ~18 months and you are the market leader. It makes zero sense when you're releasing 8 models a year
Edit: Just as I post this someone else comments google's doesnt follow elemental sizes at all so my theory is wrong too!
Here's a translation of the entire Fox and Grapes fable:
> Driven by hunger, a fox tried to reach some grapes hanging high on the vine but was unable to, although he leaped with all his strength. As he went away, the fox remarked "Oh, you aren't even ripe yet! I don't need any sour grapes." People who speak disparagingly of things that they cannot attain would do well to apply this story to themselves.
So I'm just hoping for consistency over some period of time to allow me to get familiar.
The point is that there is prior knowledge that completely explains the entire system. A 3 line poem < a 14 line poem < a book-sized work < a multi-book epic, boom done, easy. Consistent and friendly using words that everyone should know.
With google we have element 1 < element 18 < element 10 < element 6. There is no consistent schema where their namesakes map to their relative sizes, price, and capability.
So wait, which is smarter - Fable or Opus? It's clearly not obvious to most people or this debate wouldn't be happening.
As the old saying goes - if you're explaining, you're losing.
The problem with Google's is that knowledge of the periodic table doesn't seem to help.
Would love "small", "medium", "large", "extra-large", etc...
Systema would have worked. And maybe galaxias after that.
Language models being named after text/story formats make sense for anthropic, everything else is ???
Is it because Anthropic put a ton more effort into the character of the model that it basically roleplays as a human and that why it tends to think better when giving tasks that it might not be trained heavily on.
I have used Carbon more than most model to try and get the same results but it just feels like i'm talking to a smart AI. Which is fine if you are asking questions, researching an idea, but if you actually want to work on something, build something, it just falls apart. Like Opus can run for days on a task and it feels like it understand what it was doing 2 days ago even when it might not be in it context while other models feel like they breakdown and get lost or falter as the task grows.
To the point that I now use a workflow of delegating from Opus or Fable to Gemini4 to write commit messages and PR descriptions, it mostly solves the Claudese problem.
how do you do this w/o breaking bank? for Argon to write good commit messages, it must read the diffs and all the work Opus has done, which could easily eat up its usage quota.
No wonder they went for the 5.5 naming, instead of the .1 increase. From just the benchmarks, it seems like neon is basically on par with Opus 5.5 which came out 2 months ago.
Have another session on another machine (say, IOS build on a mac, and server build on a linux?). No problem, they will happily coordinate with each other through the /remote-control protocol, and then you feel like you are in a crowd, with those sessions talking to each other and over you.
No such feel with OpenAI nor Google models for me. They are absolutely great for lots of things, but nothing comes close to the feel of Opus 5.5 and Fable.
(I am biased; i mainly use the anthropic models so i got used to them).
Mad that this is seen as a long time horizon!
Other companies would use completely raw internet data for pre-training, which is basically trash. Anthropic's secret sauce is that they filtered and edited down their data for only the highest-quality material. You really don't need quadrillions of tokens to pre-train, as that's just random noise. A smaller-but-higher-quality data set is better in every way.
Meanwhile its obvious that companies like Google would use trash data like email and chat logs or Internet forums for pre-training. Just complete trash data.
This is a classic case of "less is more" and "garbage-in-equals-garbage-out"
I honestly think it just their post-training is fantastic and I think it has something to do with them being very opinionated on how they want the model to act that it basically the best at roleplaying a human. Humans are very good at completely and figuring out things that there aren't "trained" on if given basic set of intelligence. So when it working on these long tasks that might not be directly in it training data it still reason through it very well.
They got made fun of for this in the pass because they do a lot of research on this including hiring a philosopher.
This is just my best guess, because you don't see any other labs talk about how they made the model act and Anthropic is basically obsessed with it and talks about it every model release.
I think they just have better "taste" and that taste is their secret sauce.
Everyone uses books as well as internet data.
Anthropics difference is they have a huge data budget for buying RL environments for real world tasks.
If you look back at GPT2 and the early days of AI safety/risk this was actually something the kinds of AI researchers that joined Anthropic didn't want - RL on coding->using that accelerate development and to bootstrap capabilities across domains was always the plan, not making it easier for people to spam/fake homework or trap people in problematic human-AI attention/relationship tarpits.
I personally find that Opus 5.5 and Fable 5+ are generally fine to chat with but just very wordy and culturally "LessWrong/essay-guy". Which, if a model is going to have a personality, is close to the one that I want (though it does piss me off when it gets preachy about standards Anthropic clearly doesn't hold themselves against)
I also want to immediately recognize AI generated content on the web or from another person, so I'm very glad it has a distinct voice
From simple frontend work to strategically adding virtual rulers when working with LIDAR scans to extract more accurate data, there are very clear points when Anthropic introduces a new external data set into their mix and while I have seen the same from competitors such as Google, there simply appears to be a greater tendency to stick with these once they have them.
As an example, at release it was quite apparent that Gemini 3 Pro, 3 Flash and Opus 4.5 had a similar, possible even the same frontend-focused training base bought in from a third party. But while Anthropic has had both smaller and larger improvements across each subsequent Opus release, I have not noticed such a consistent improvement with Gemini models over the last year. Maybe Anthropic simply buys updates more often, maybe they create more artificial PRs on top, who knows.
Same with the LIDAR scans, there was a point a few months ago when models started to exhibit a very systematic way of handling visual references in some few leather-crafting tasks, enabling whole new ways of working more efficiently combining existing measurements with images. Step-by-step pixel measurements, compensation for angles and poor captures, etc. Have since seen this in other models, again very much feels like a bought in data set that they are pushing as far as it goes.
Then there is 3D capability with models like Muse Wonder Full. I have not used it myself, if it ain't on API, it doesn't exist, but have seen PRs and do find the model output very parsable, despite it being terrible at any modelling task because these capabilities do not emerge from just training on vast amounts of data, they require high-quality, well tailored, step-by-step reference data of the entire process.
Harness wise, while pure output performance is very consistent across harnesses in my evals (I nowadays use both OpenCode via API and the labs own harness three times each to normalise), Anthropic also appears to be far more willing to use the time between releases to prepare features for the next. Stuff like workflows, proper subagent implementation, advisor, etc. often have found there way into Claude Code either in-between models or were first seen with a model that couldn't fully leverage these. Also feel that explained the good overall launch experience Fable 5 delivered, they simply used the time between announcing Glasswing and launching full access to get the UX/DX polished.
Meanwhile, I am still testing GPT-6.2 Astra and Sol so will refrain from making any more judgement on either till I have all data, but apparently there has been an extensive last minute rewriting of the system prompt to improve performance and "behaviour". Given the gap between GPT-6 Astra and yesterday, along with statements by OpenAI staff alluding that they had this model ready for a while, such things should happen in a far more structured, systematic, well thought out manner, not last second after some negative outside feedback. Heck, from what I know Anthropic, unlike any other lab, does zero early access outside the limited hardening for very exclusive partners like Glasswing. Certainly no early access for regular users, reviewers or the like from what I've heard, so Anthropic does all that prep-work for model releases with less feedback than other labs.
I'm planning to write a blog post on it, but my theory is that it's aligned to not be "politically correct" but instead to have a solid stance. In the Claude Constitution, they explicitly say how they train it to have judgement and not "follow rules"
Cursor was popular before any of those three, and just look at how much Grok improved after they acquired Cursor's data.
No one that I know of uses Cursor anymore, and Grok has been disappointing since ages.
So they can train on prod traces or at the very least use user prompts to seed good synthetic tasks to train on.
I doubt there's a single reason for why Anthropic are ahead right now. Building a frontier model requires executing well in a wide variety of things. It's probably a result of being slightly ahead in various aspects of of data collection, model architecture, pre and post training, harness development, etc.
It's a shame that so much of what's happening is being closed doors. Maybe one day, decades from now, somebody (or some AI) will write a definitive history of this period, explaining what was going on at the various labs, how their research compared, and what the major players really thought.
I'd recommend trying Gemini Neon on Antigravity, with a trimmed down (or new) agent configs.
(After a long conversation with existing messages)
User Prompt: Ok, so SSH to the unraid server and update the plex db
Reply:
I need to think how to answer this.
The user is asking me to SSH into the unraid server. This seems to imply there is a machine to which we can SSH to. I should look for places where hosts may exist.
The implication is that once SSH is established to unraid, the plex db can be updated. It seems plex is a service that the machine has in which we can run updates.
I'm wondering if plex might be the Plex service the user has been discussing.
So the user doesn't want to SSH to the server - actually they want to SSH to the server.
... Formulating a plan ...
[some messages later]
Aha! THIS IS THE SMOKING GUN! `~/.ssh/config` HAS THE REQUESTED HOST!!! IT'S TIME TO CONNECT!!
[some messages later]
Ok, time to act. I'll ssh to the server and follow coherent rules.
----
I am shocked at how primitive its reasoning sounds (it reads just like gemini 2.5-pro's old reasoning). It often gets the job done (although it's quite slow and can't tackle more sophisticated problems).
Perhaps Anthropic's internal reasoning is also like this, but they really rewrite it to make it seem more coherent (which certainly sells it on me)? I almost want to click stop just by reading this nonsense.
I was not really fabricating the example I gave too much. Here's a verbatim transcript from a session which was filled with this stuff:
It seems rather ironic that Anthropic givens the most logical reasoning traces, though. You'd think if they're ahead they'd either hide them or make them produce garbage (like what seems to come out of Google's models).
If you want to see example traces without jailbreaking claude or gpt, see open models that display the CoT without trying to hide it. It's pretty useful sometimes, with usual caveats (it deceptively looks like English but is heavily reward hacked, and it's generally a weak proxy for model's reasoning process).
You're of course completely right. Interestingly the "summaries" have only become more diverged over time. I now believe the big 3 are barely summarizing at all any more. I deem it likely that the small models don't get the CoT at all, they just completely make up something "plausible" solely based on the user message with a little bit of previous context. Much of the time it will be the complete opposite of something I asked for, or barely related.
Anthropic has changed this up multiple times over the last 12m, I'm talking about the iteration that was around the longest this year. I think they changed it again within the last feww weeks and potentially started actually summarizing again.
Your suggestion is the right one. Everyone who uses LLMs for their job should spend some time with open models like Deepseek and see what the thinking actually looks like.
(And it did produce a very competent algorithm, though it took a few days and a metric ton of tokens; I just triggered the same task with Neon).
Just today I was filling out a security questionnaire and it said "I won’t write a PDF that claims zero credentials while this token is in the logs." But we were in the process of removing this from the logs so it was perfectly accurate to write and I wish it had deferred to my explicit request...
I think Anthropic overcorrected on the sycophancy; Google models to me seem a nice balance of highlighting issues / risks while still doing what I actually ask.
It might be that opus is better in better structured projects, and carbon is better at (re)structuring poorly structured projects?
Carbon is a better project manager/product owner - opus a better programmer?
Lots of PR it seems?
I have all of Ant/OAI/OpencodeGo subs and test lots of models and I just cannot understand how and why some people find specific models so ahead, when it seems to me like reality is so nuanced and so specific to the particular task and how you approach it.
Google is the best example of this with their product graveyard. No matter what they release- they just can’t make a good/nice/usable product.
Just check google ai studio as one of latest examples.
With the frontier getting more jagged, we might see more of this 'less intelligent overall but useful for the frontier tasks'-type models, meaning picking a model will only get more difficult from here.
(FD: I work at AA)
I do apologize in advance for putting you in spot like this, but there are two things I have been quite curious about.
1. Since 60% of the benchmarks contributing to AA are now private only, how do you guys ensure transparency? More precisely, how can someone verify the transparency in scoring?
2. There are also disclosures about custom benchmarking services offered to AI labs, what does AA do to keep the custom benchmarking services not getting access to the same/similar data that is used in the private benchmarks?
I looked into this using data I have from a battle system built on top of OpenCode (so we're trying to capture real world use). It looks like Carbon actually does a lot better when it uses significantly more output than opponents.
Win rate when output is less than 2.5x opponents: 53% (41/78)
Win rate when output is more than 2.5x opponents: 73% (41/56)
The p-val on this is 0.02. Data from https://agentduel.com/leaderboard.
There doesn't seem to be any relationship between actual size of task though. Win rate appears random based on total output of Carbon and also opponents. (Though our data is still small, so it might exist.)
Looks basically like Carbon was trained to think a lot longer, not necessarily that it's better on large tasks.
I’m out of my depth here but it does seem prompts have a “smell” to their complexity that means we can use a classifier to mate them to a model with better success than random noise, but still don’t essentially all of the savings associated with systems like this just come because there is some “complexity arbitrage” where there is enough of a token cost differential between models that you always save money by trying the cheap model first and then trying a hard model if it turns out to be hard?
What I want is a deeply introspective model that knows to take a step back and call in for aid.
It's really frustrating to describe. Easiest is you can't score on just the prompt, you need the full context. Ex. "Solve Fermat's last theorem" initial prompt then "continue" followup prompt. Which gets you back in the tarpit of "if I have to pay a model to do an inference run on context, then, the context is cached and I should continue using that model."
The vision of another model thats RL maxx'd on knowing how to delegate has a few problems. One, we can observe externally, it died/didn't work even at OpenAI's scale and willingness to make it work, i.e. the...GPT-5.0? IIRC...router that silently went away. The other, if such a model was built, it'd also be good at regulating it's own thinking enough that it'd naturally use less tokens.
In practice, the frontier hackers have settled on a decision model (i.e. Jev) x calibrating effort level or choosing from a set of models x reasoning effort tuples they specify rules for choosing in their Jev prompt. My guess is this stays stable and the only real choice.
I'm not sure it's much better than 4.1 Flash vs 4 Pro -> if 4.1 Flash is better temporarily when Pro is typically significantly better, people are going to be confused.
I'm not really sure how you get around this problem easily.
It's not going to age well when the next Flash-lite model launches, but maybe they will call that `Gemini 4 <Element>-lite` - who knows...
I've found that while Opus 5.5 is often pretty nice for practical work, Fable (even 5.0) is still significantly smarter on hard tasks and makes far fewer "dumb mistakes"
That would make a lot of sense from an inference optimisation perspective. The same prefill and decode machines can hold one set of weights, and serve both Astra and Sol requests.
At (multi-)terabyte-scale model weights, and when you have to offer not just the latest like Fable 5.2 but still serve 5.1, 5.0, etc; heck I think Opus 4.6 is still around; you really care about overall compute utilisation.
Idle GPUs are wasted GPUs; and even over the fastest interlinks you’re not hotswapping multi terabytes of weights.
With MoEs, there’s probably stuff you can play around with varying the number of active experts; or other clever optimisations.
Maybe Sonnet 5.5 could be just Opus 5.5 with a different router and less active params? Who knows; I’m just theorycrafting now :)
This doesn't really make sense given how the models released.
OpenAI released Astra long before latent-space looping was a thing at the frontier space.
Then they released Sol 6, which was fast, clearly not looped. Then they release Sol 6.1 which was nearly as good as Astra, but way cheaper, and WAY slower -> this would follow it just being a looped version of Sol 6.
It doesn't seem likely that Astra 6.2 is just a looped Astra 6... It's too fast, and they'd go bankrupt if they're charging the same price... It may be a further looped version of Sol than Sol 6.2 - which would explain why it's not as much better as we would've hoped.
It may also explain why their release is not going as well as they hoped. It's hard to predict how your first looped models are going to fare.
If you look at how long it took Google to recover from their failed 3.5 Pro, it would be surprising if GPT recovered from their rumored failed run on Astra in under a month. That would give them reason to release a looped Sol and call it Astra.
Even in the open weight Chinese models as small as DeepSeek V4.1 Flash - nothing is looped.
I think the rumor-mill consensus is that Google's models are significantly smaller while attempting to swing at the same level - but the ratios AFAIK are thought to be around the same - i.e. Neon is somewhere around half as big as Carbon, big enough that a newer checkpoint can make it temporarily a little bit better.
A GB300 rack has 72x288GB=20TB VRAM. Do the math and 10T parameter models are just about the max size you can easily train and serve (at 1M context) without doing a lot of custom hardware development or limiting your scale.
https://news.ycombinator.com/item?id=19393432
(Although why they'd release all ~four is a mystery. I guess they're having a hard time picking the "best" one and so are just putting them all out to let the public decide.)
It doesn’t correlate with intelligence, speed or price furthermore you have no way of telling what the model sizes of “like for like” named models from oai or Anthropic are.
> In our early evaluations, it matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost. We’re excited to bring these capabilities to Glean customers tackling multi-step work that demands accuracy, detail, and speed.
So possibly smaller than Carbon? Or some other axis that is making it faster and cheaper? No idea.
- Gemini (New)
- Gemini
- Gemini (original)
- Gemini (Not ready yet)
- Gemini (Deprecated)
When Argon first came out it seemed like they could be going for noble gasses, like Helium (small), Neon (mid), Argon (big).
But then Carbon came out and it seemed like maybe they were going through the alphabet for model generations. Argon, Boron (internal), Carbon.
And now they've come out with Hydrogen and Neon, and there's no logical way to order them. The actual model sizes are Hydrogen (small), Neon (mid), Carbon (big). This doesn't match sorting by atomic number or alphabetically.
I think there are a ton of people who automatically default to the largest model with the highest thinking for everything. Which historically made sense but not nearly as much anymore.
At this point a model like hydrogen can cover just about everything an average person is doing. You don't need Carbon deploying an agent swarm pushing terabytes of parameters through HBM memory to help you plan your birthday party.
Worse, people burning dumb prompts on large models also has the net effect of shrinking quotas for everyone.
(M) Gemini 4 Neon - second fastest - $3/$15
(L) Gemini 4 Argon - second slowest - $2/$10 (till Nov 30), $4/$20 (> Nov 30)
(XL) Gemini 4 Carbon - slowest - $2/10 (till Nov 30), $4/$20 (> Nov 30)
There is a Helix model too (in the pipeline) which is the actual small size.
element/model - atomic number - cost H - 1 - $1.50 & 7.50 per mtok C - 6 - $4 & 20 per mtok Ne - 10 - $3 & 15 per mtok Ar - 16 - $2 & 10 per mtok
Claude - length of story but fable messes it up Haiku - 3 lines Sonnet - 14 lines Opus - a novel? Fable - 30 lines?? Mythos 15700 lines?
Luna - 28 days Sol - 365 days Astra - 5-6000 stars
I also like the personalities implied by these, with Anthropic saying “human literary achievement is neat”, OpenAI saying “space is neat” and Google being like “science is neat”.
Edit: I was coming back to say “I wish we could talk about stuff more substantive than the naming scheme” but (1) my comment may be objectively incorrect, and (2) I have to apologize that I haven’t helped the situation by… commenting on the naming scheme.
So, names don't align with capability but instead "release freshness". This is unlike every other provider but it is what Google does (did?) with Android releases I believe (dessert names).
But the way google ranks the models by capability doesn't follow periodic table order.
Hydrogen - 1
Carbon - 6
Neon - 10
Argon - 18
It could very well be based on their density (g/cm^3 in standard conditions):
Hydrogen - 0.0000899
Neon - 0.000900
Argon - 0.001784
Carbon - 2.267
This would make more sense than assuming it is based on their periodic table order as it more accurately maps to how Google markets their capabilities, but is too esoteric to really be useful for clearing up confusion for normal users, IMO.
Carbon has a huge problem that it's just way too slow, and outputs way too much per task (on real world tasks, it's the worst at this https://micjm.com/blog/carbon-is-that-slow/).
Without the speed issue, Carbon would be a top model already. It's top on our battle system built on opencode, if you exclude early votes: https://agentduel.com/leaderboard?excludeEarlyVotes=true, though to be fair, small sample size so far.
Conclusion is that it's looking good for Neon. Even if it suffers from similar problems as Carbon, if it's even smarter and way faster, it should be a top model, maybe the best.
Edit: I was wrong about Neon being the larger model. I made this assumption because Neon has higher scores and they're both "v4".
Carbon is a Fable sized model
Carbon seems to be a version of Argon that is incredibly spendy with thinking tokens.
Weird because Neon is outscoring Carbon in AA index, Terminal bench science, etc and they're both named "4". Should have released this as Neon 4.1.
With that clarification, I'm sure Neon has the same overthinking problem.
Google also has a rumored Krypton internal model, which is also Argon/Carbon class. Neon might very well be distilled from Krypton, not Carbon.
> Neon might very well be distilled from Krypton, not Carbon.
That would make a lot of sense, interesting!
Also neon seems to be stepping on the toes of Carbon a bit, just like Opus 5.5 did to Fable.
This is probably the set of benchmarks most people care about: https://www.vals.ai/comparison?m=claude-fable-5-2,cobalt-c-h...
Overall, performs just under Fable (above Astra or Sol 6.2 or Carbon - slightly ahead of Opus 5.5).
Latency is by far the best for the level of quality.
It's about 2x more expensive than Sol 6.2, but significantly better and faster. You get what you pay for.
If I wanted to realistically try out Neon in some of my workflows without committing to a paid subscription before seeing if it can actually work as well as Opus, do I have any option, or do I just need to take the gamble?
"Using third-party software, tools, or services to access Antigravity is a violation of our Terms of Service"
and also: https://antigravity.google/docs/faq?can-i-use-antigravity-cl...
I'd like to keep my Gmail account, so Google can keep their model and will not get my business.
That being said, you can do whatever you want with API tokens.
Paying for Gemini at API rate is therefore out of the question for me.
IIUC, you also don't get to try Opus unless you have a subscription, right?
Edit: I tested the model and it found several mistakes Astra 6.2 made. It’s possible…
Unfortunately, their recent change to allow me to opt out of AI training (which would also include connector data like Gmail, photos, etc) hasn’t rolled out to me yet[1]; so I won’t be upgrading my subscription; nor using the Gemini app until I can opt out.
[1]: https://arstechnica.com/ai/2026/11/google-will-let-you-opt-o...
[1]: https://blog.google/innovation-and-ai/products/gemini-app/up...
Which ... doesn't necessarily show what you want either. But you can access "remote history" in bottom left. Maybe, if it works for you.
So a regression then.
/s not trying to be mean, its just weird hill to die on. Every company is stealing all your data in ways you unknowingly consent to at every second of the day.
The latter need to raise funds and IPO, but the former is an arm of one of the most powerful companies to ever exist with some of the largest data warehouses around and custom hardware. And they have an endless pile of internal code to maintain already with hundreds of millions of end users already bought-in.
The incentives being different leads to different focuses and product lifecycles.
Costs spread between Hydrogen ($6) < Carbon ($11) < Argon ($13) < Neon ($17) for this 43 question exam.
At a quick glance, the tone is good as well. But we do a fair bit of prompting in our harness to guide the model, so I can't say how the default model behaves.
[1] https://plotly.com/blog/gemini-4-neon-plotly-data-analytics-...
What I simply don't get is why people want to push the boundaries of efficiency and knowledge. I used to think as a teenager that all knowledge was good but how good is knowledge when we can't acquire wisdom with a comparable speed?
The greatest thing about the early internet and earlier science was that people shared and communicated with each other. Now they just share what they produce wihouth much humanity at all.
It seems to me that the creation of all these models has just become a race where no one really cares about the finish line and no one really understands why we're doing it, except for the short-term trail of dollar bills on the race course.
For me, each new model release is a sad day for humanity.
ChatGPT thinking: the user is not happy with my previous solution, propose a cooler scheme like toy story character names...
This is an underrated quality of Gemini models.
https://artificialanalysis.ai/#intelligence-efficiency-tabs
Race is getting very interesting again.
Google Antigravity
Entry rate limits to advanced models, including Gemini 4 Argon and Opus 5.5, in Google Antigravity, our agentic development platform
Yet, here I am, running the `agy` cli, and the model picker says "Unavailable: Gemini 4 Carbon/Argon/Hydrogen/Neon is currently available to paid Pro and Ultra subscribers". It has been this way since the release of Carbon. I guess it's "3.8 Flash" or nothing for me.
Gemini 4 Argin vs 6.2 Sol vs 6.2 Astra vs Fable 5.2: https://aibenchy.com/compare/google-gemini-4-neon-xhigh/open...
Eventually all these models will have to merge into a Omni model with image/video/tts voice/text all in one.
You try keeping 600+ models in our tables updated from all the model providers; it's a part-time job every day.
On Artificial Analysis Intelligence, it scores comparably to GPT-6.2 Sol but is significantly more expensive.
https://bitbin.de/ramen-bench/#google%2Fgemini-4-neon%2Fxhig...
They all need to pick a dam naming scheme and stay with it.
So Neon is the new default, and Carbon the slow deep thinking “when you need it”? A bit like how Opus 5.5 was better than Fable 5.1 in most things, but Fable still had the edge over deeper things (then 5.2 came along)
Hydrogen has already became my default for 80% of small tasks. If Neon is as fast but smarter that would honestly be very cool
So I'd take their hierarchy with a grain of Sodium.
That spread is my personal bs detector for models over promising.
But I still can't understand the differences between these models.
Recently I watched Veritasium talking about helium, and I was thinking, will the effort setting of these models eventually have something like "Temperature K" to configure?
https://x.com/andonlabs/status/2123174738789011875
> We introduce Riemann-Bench, a private benchmark of expert-curated problems designed to evaluate AI systems on research-level mathematics that goes far beyond the olympiad frontier. Problems are authored by Ivy League mathematics professors, graduate students, and PhD-holding IMO medalists, and routinely took their authors weeks to solve independently. Each problem undergoes double-blind verification by two independent domain experts who must solve the problem from scratch, and yields a unique, closed-form solution assessed by programmatic verifiers.
https://arxiv.org/abs/2604.06802
and why can't I use any of these with agy and a $20 plan? unlike openai and anthropic
Edit: I stand corrected, it can be used in agy.
Edit2: I have an student account and students accounts can't use these models apparently, https://discuss.ai.google.dev/t/bug-entitlement-sync-paid-go...
I'm curious about Neon.
Maybe the people at Google don't realize that but some people need constancy. If a model was good enough for some work 6 months ago, it's not suddenly becoming not good enough for the same task.
So Carbon is already deprecated, or?
EDIT: They do explain where it sits in the line-up here:
https://blog.google/innovation-and-ai/models-and-research/ge...
So now we have
(XL): Carbon/Argon Astra Fable
(L): Neon Sol Opus
(M): Hydrogen Terra Sonnet
(S): Helix(rumored) Luna Haiku
Also, very curious whether they will ever release the rumored Krypton model, or will it be hidden ala Mythos
[0]: https://idiallo.com/byte-size/buying-ai-for-your-company
On Google's scale you could probably mint a datacenter of free compute if you could get people to stop using 10T parameter giants to vibe check their emails.
Here are the others in the family: https://static.simonwillison.net/static/2026/gemini-4-pelica...
Good lord, this is an insane flow.
I do think they're very well positioned, their small models are smart and fast, and integrating well with Google is a big plus for basic use case stuff. I always like these models!
Nevermind the 20GBP pcm I pay for YouTube.
Can't wait to buy Puts on ANT...
And then you have big names like Hy-dro-gen haija
I am confused with the naming scheme as well, and it's a fun topic, right?
2027 january RSI is achieved. 2027 february first big scary incident (100 million people severely affected by infrastucture collapse for several days) 2027 march a fleet of rogue self replicating agents is discovered. 2027 april first severe fatal incident (1000+ people dead, most likely from a missile created using oss models). 2027 june. Deadly virus created by a terrorist organization is released in Ukraine. Remind me!
Maybe the atomic number correlates to model pricing? Nope! Neon (10) is marketed as cheaper than Carbon (14).
Maybe the atomic number correlates to overall intelligence? Nope! Neon (10) and Carbon (14) are both marketed as smarter than Argon (18).
The only thing I can think of right now is that it's in alphabetical order by release... but in that case why the periodic table?
haiku -> hydrogen
opus -> neon
fable -> carbon
No mention of Argon in that page though.
e: No, Neon is the cheaper, mid-tier between Hydrogen and Argon. So Argon/Neon are aligned along the noble gasses, and Hydrogen kinda makes sense as the weakest ... but they threw Carbon in without any consideration
Just give them a version number and an identifier, like:
1. V5 smol 2. V5 mid 3. V5 big 4. V5 yuge
WTF are Hydrogen, Oxygen, and Nitrogen? At least OpenAI and Ant have Luna, Terra, Sol, Astra and Haiku, Sonnet, Opus, Fable. They use those consistently.
Say what you like about Apples Pro, Max and Pro Max -- at least it gives some indication of size to the average person.
Luna->Terra->Sol->Astra is fine, you need to be familiar with meaning of the names but it makes sense and works.
Haiku->Sonnet->Opus->Fable/Mythos is better, more common words, more obvious understanding.
Hydrogen->Argon->Neon->Carbon (is there a Barium too somewhere? ) fck all I have no idea, people don’t even know noble gases well, but hydrogen and carbon doesn’t even fit that. 1,2,3,4 would be a better naming scheme.
That sounds like serious benchmaxxing for Gemini 4 Neon is ahead on 17 out of 19 benchmarks vs Opus 5.5.
Not so bad for a company that, according to HN, "has lost the AI race".
P.S: I like the Google models but a truce before the Anthropic IPO would have been nice. This is just a very big dick move a few hours ahead of Anthropic's IPO. Shall we see Anthropic launch a new model from now 'till the opening bell?
Gemini is not very helpful there. It just kind of gave up.