1037 points | 3d ago | Discuss on Hacker News | Back to Radar
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
The real question is: will it be available for Google Workspace? Not everyone wants to pay for a dedicated AI license :-(
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...
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...
I heard that Google share price doesn't really depend on how much their models are used. But rather how do the bench and whether they are theoretically publicly accessible so that their quality can be independantly confirmed. So Google can price their model higher to get most of the benefits without having to actually provide too much inference.
They are probably releasing the models to the public at all only because they got so much backlash after witholding Argon.
Gemini 4 Neon (NEW!)
Gemini 4 Hydrogen (NEW!)
Gemini 4 Carbon (NEW!)
Gemini 4 Argon (NEW!
Gemini 3.8 Live
Gemini 3.8 Flash (NEW!)
Gemini 3.7 Flash
Gemini 3.6 Flash
Gemini 3.5 Flash
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. Hydrogen 1
Carbon 6
Neon 10
Argon 18
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.
It's a good model for delegating work to, but not deciding the detailed scope. So, not fable class for me.
- Gemini (New)
- Gemini
- Gemini (original)
Hydrogen is the cheap one for straightforward tasks.
Neon is supposedly Opus-level, which is the exciting one if it actually works as well as they claim.
Obviously anything with a 3.x is not what you should reach for. This has always been the case on a first digit bump in any of these mainstream models.
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.
Surprised if it’s not going to be 4 electron/proton.
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.
(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.
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 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.
Neon and Argon are descending down the noble gases, I'd expect their successors are Krypton, Xenon, Radon.
They skipped Helium and went straight to Neon, which makes me think they've reserved Helium as a lightweight element, and are going to follow the lightest elements for the cheapest models. Hydrogen, Helium, Lithium, and so on.
Carbon reads more as a "foundation of life" kind of metaphorical choice, and they have some choices in the periodic groups nearby to keep that sort of metaphor for a while. My guess is they move horizontally - Nitrogen, Oxygen - then on to Silicon when they hit something appropriately transformative with the model.
I'd expect the most experimental or fastest models to eventually release on highly reactive groups - Fluorine, Sodium, etc. - and that they're leaving the metallic elements in the middle for differentiation in industry-specific fine-tuned models.
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
(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.)
and why can't I use any of these with agy and a $20 plan? unlike openai and anthropic
On Artificial Analysis Intelligence, it scores comparably to GPT-6.2 Sol but is significantly more expensive.
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.
Edit: I tested the model and it found several mistakes Astra 6.2 made. It’s possible…
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?
IIUC, you also don't get to try Opus unless you have a subscription, right?
"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.
I never had any issue using Claude for code review in Codex and vice versa.
I remember that half a year ago, Anthropic planned a change to make claude -p draw from a separate usage pool at API pricing, but after feedback that change was scrapped.
What people were banned for is using the subscription authentication in a third party tool, using many accounts, or reselling.
Invoking the official CLI from another agent for legitimate personal use during development should not be a problem.
The requirement is that you use antigravity harness to access subscription-based accounts, not that you can't wrap it
> Wrapping agy inside third-party coding agents or applications, or automating account rotation, is not supported under consumer plans.
At least they no longer ban the whole Google account, so I think interpreting that as a prohibition only for abuse would be reasonable.
That spread is my personal bs detector for models over promising.
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
01 02 ... 13 14 15 16 17 18
H He
Li Be ... B C N O F Ne
Na Mg ... Al Si P S Cl Ar
???It's not like this has no precedence. Google themselves are the best known example in software with Android codenames. Starting with [A]rgon was a clear signal in and of itself. But above all, just basic common sense strongly goes against Google uses some very limiting, niche pattern rather than just alphabetic.
I'd chalk it up to eternal September but I'm not convinced most of these people are new. A particular brand of neurodiversity, I guess.
1: is this use of "dense" obsolete given dense LLMs are often smarter than sparse models?
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.
Which of these are missing in your opinion? I'd argue we are there already. At least very very close, as "with maximal efficiency" is quite an arbitrary term.
The growth mindset isn't really for you, if you're already rich, and don't want to be richer. It's coming from the vast majority of the world, which is poor, and the minority that is rich but wants to get richer.
We don't want a US cost of living, we want something like a Western European quality of living, and we want to figure out how to get that across the world even for places that have much higher populations than Western Europe, and at a low environmental impact.
Comparing how much a decent quality of life currently costs in the most expensive countries isn't useful.
Also, we need to focus on the basics rather than just trying to give people tons of money: good food, decent housing, free healthcare and free education.
Do you think "~$4,000 to ~$12,000 at a German cost of living" would have made it clearer how much we actually have to go around today? Or maybe just saying "the global economy isn't even close to the size where we produce enough for everyone to have decent food, housing, healthcare and education"? To me that seems likely to interpreted as vague and easily dismissed.
A major point will be complete training-data transparency with independent expert review. In EU a statutory ban on collecting/storing/using users' data for training (including the "de-identified data" loophole), except voluntary explicit opt-in.
I wonder if a mandatory disclosure of training recipes paragraph will be included, likely in EU only (e.g. with tiered-access design to counter trade-secret objections) ?
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?
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...
/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.
Which ... doesn't necessarily show what you want either. But you can access "remote history" in bottom left. Maybe, if it works for you.
[1]: https://blog.google/innovation-and-ai/products/gemini-app/up...
So a regression then.
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.
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.
Say what you like about Apples Pro, Max and Pro Max -- at least it gives some indication of size to the average person.
They all need to pick a dam naming scheme and stay with it.
Can't wait to buy Puts on ANT...
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.
(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.
(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).
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).
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).
Sometimes I'm like...is everybody else using the same thing that I am? This is insane.
They definitely let it churn until it figures something out, it's not uncommon for it to take 15-20 minutes to answer, is that normal? They just have "relentlessness" maxed out and some kind of second-order structure to orchestrate a bunch of agents into one unit of work.
It's obvious that what we're getting are summaries. Sometimes I think I even see vestiges of the summarizing model instructions. But, that being the case, it's bizarre why Gemini's summarizer acts like a toddler, and Claude's actually lets me learn and, more importantly, stop it at its tracks when I see it's going down the wrong path.
On the other end of the spectrum, if you toy around with the open-weights models, you see the insanity that is going on in their output. I can totally believe in a scenario where claude is basically doing that and then summarizing it into something coherent, getting rid of the 10 opposing viewpoints, "aha!" and "Hmmm, but wait!", but it does it elegantly.
I've tried constantly to use other models (and I do), but Claude spoils me to no end. I keep thinking I must be holding everything wrong.
Yeah, the open weight models (the only ones that show the reasoning tokens) are just absurd. Wait but why oh wait now I have it, this makes total sense!
I find it kindof endearing, tbh as I definitely have similar trains of thought when developing.
This is like, the entire ballgame. This is what effective communication is, this is what understanding is. Anyone who can do this, man or machine, would do so, I would think.
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.
I'm not sure DeepSeek 4.x is a good example; its reasoning traces are extremely verbose and detailed. Which is probably one of the reasons it performs so well, and appears to be a deliberate architectural decision to improve skill at the cost of token usage and time-to-answer.
A better comparison might be Gemma or Qwen. Running these locally and with no summarisation, the reasoning traces are not all that different to the big proprietary models in terms of verbosity and content, even if they're not as smart.
That's why I recommended it. It shows that for a given model, it's CoT might look different from what people expect. I also named it because I saw in another comment here - or maybe it was on a thread about Carbon, one of the two - chatter about Gemini 4 outputting reasoning inside its user-facing message and looking like "I'm not sure. Maybe it's A? YES! Now I get it! ...", which is of course the style that Deepseek 4.x's traces also have.
I haven't even really tried to use other models in the same way, and this thread makes me feel like I don't need to.
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).
I had subscription to both anthropic and openai. I was out of credits for openai, so used anthropic to fix some coding issues. Spent whole weekend prompting while it kept failing over and over. Finally on Monday I had openai's weekly quota start, 2 hours issue was done.
The same thing happened again, after a couple weeks. Two weekends wasted trying to make Anthropic models to behave.
If model versions are anything to go by, I'd have assumed Opus 5.5 and 5 were built on the same pretrained core, but the vastly superior performance of 5.5 makes me wonder if Anthropic literally pretrained a new one using Fable (following their RSI advances).
Opus 5.5 may be the first RSI'd model and it shows.
Mad that this is seen as a long time horizon!
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.
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.
Btw, just one B200 x 8 server is all any mid-size company needs: any open source harness + any 80 - 120B active params model - it will fly and serve hundreds of agents. All data and processes are yours - no subscriptions, full control.
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.
I never used Grok before the Cursor acquisition (we use Cursor mostly), and if this is improved, I shudder to think what it must have been like before the acquisition.
It's like an irritating Sonnet 3.5.
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"
In my opinion, the Argon class models (Gemini 4 "pro", so not Hydrogen). have the best writing voice.
It's still LLM-ese mind. But it's easier (possible?) to prompt it into some type of submission. Claude is very resistant to prompts telling it (how) to write better.
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
Discussion about issues with Google LLMs: "Have you considered that Google are simply bad and failed to do that properly?" (saw this exact line at least a dozen times now on HN)
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.
Given all the other things Google should be good at and is failing at why would this be the exception?
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.
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.https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....
They were not trained on good data.
And higher-quality source material results in higher-quality emergent behavior, everything from thinking to tool use.
Seriously people, read their copyright court case. It literally states why Anthropic focused on books and higher quality material.
Google has to learn to not use trash internet content if they want to catch up.
(note that I haven't tried Neon yet so maybe they learned their lesson and fixed it. All my experience is from 3 months on their Ultra plan and previous models from earlier this year)
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.
> According to The Information, leaders at Anthropic have discussed spending more than $1 billion on RL environments over the next year.
https://techcrunch.com/2025/09/21/silicon-valley-bets-big-on...
OpenAI: https://news.bloomberglaw.com/ip-law/openai-risks-billions-a...
Google: https://www.theguardian.com/books/2026/jul/14/publishers-sue...
There is zero evidence that book data significantly increased coding performance. It increases coverage of knowledge, but not much else.
"Over time, Anthropic came to value most highly for its data mixes books like the ones Authors had written, and it valued them because of the creative expressions they contained. Claude’s customers wanted Claude to write as accurately and as compellingly as Authors. So, it was best to train the LLMs underlying Claude on works just like the ones Authors had written, with well-curated facts, well-organized analyses, and captivating fictional narratives — above all with “good writing” of the kind “an editor would approve of” (Opp. Exh. 3 at -03433). Anthropic could have trained its LLMs without using such books or any books at all. That would have required spending more on, say, staff writers to create competing exemplars of good writing, engineers to revise bad exemplars into better ones, energy bills to power more rounds of training and fine-tuning, and so on. Having canonical texts to draw upon helped"
From: https://copyrightalliance.org/wp-content/uploads/2025/06/Bar...
genome sequences are probably very dense in “reasoning”
I'd recommend trying Gemini Neon on Antigravity, with a trimmed down (or new) agent configs.
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.
Second: you don’t need to waste a lot of tokens having the other model re-read commits and code. My AGENTS.md/CLAUDE.md specifies to write PR summaries and titles and then hand them off to Gemini to clean them up. Very occasionally this requires two trips if Gemini changes something the other model disagrees with, but usually results in very clear and readable PRs, often better than I would have written if I spent 15 minutes on it. And it uses very few tokens.
> Authorable rules inspect the graph for structural tells. Every finding names what matched and the exact text that triggered it.
(I don’t know if it’s just me but those are also very much garden-path sentences for me.)
I personally hope this never gets solved so the valuable signal of “the person sharing this thing cared enough to describe it themselves” remains detectable.
My fear is that folks who read Claudish all day everyday will start picking up its style.
I would bet you still get tells like overusing "never" or ubiquitous inanimate subjects with animate verbs ("questions that decide the option").
For planning documents about complex work, Hydrogen does a fantastic job of both the analysis and the writing.
Who's reading that slop? Humans?
im gonna steal that
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.
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.
That is a massive, massive compliment. You literally need a phd there to set up billing so they can take your money.
Vibe coders make more usable products. There is absolutely no excuse or reason to put products of that quality out in the open. But google keeps doing it over and over.
Which just proves - you can't change company culture. So new startups always have a chance as incumbents just can't change due to "laws of physics".
>> what is a better lightweight studio product than AI studio? A vibe-coded clone is better.
For instance:
https://x.com/SPAC89/status/2123171830739632189?a=WIEEh4US46...
https://x.com/chetaslua/status/2122705177203642485?a=7aENHW-... (woops this was carbon, but still cool so I'll keep link)
https://x.com/kimmonismus/status/2123168299890704735?a=IWO2O...
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?
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"
Secondly, they somehow made all this work on public clouds with really crappy (compared to baremetal nvidia) networking. I think thats probably a lot of post training (which doesn't need such a large/tightly connected cluster), and some architecture that is less reliant on a tightly connected cluster for pretraining. This would also explain why they have been able to use Trainium, and somehow beat Google while using TPU's.
They both seem to break down at similar size codebases or complexity levels, with maybe Opus having a slight edge over Argon.
For casual use it's the harness for me, using Gemini's web interface feels like I'm fighting with all the options while on Claude's I just shoot and let the AI activates whatever tooling they need
It's possible that after 4o OpenAI has RL to make its models less human.
Maybe someone used all with API and goose or open code can chime in here.
If they've done anything with their special sauce, it appears to be to train it on the behaviours of jerk engineers who are smarmy, overconfident, and arrogant.
I prefer to use the Gemini models - if I need to spend 8 hours a day talking to something, I dont want that something to be sanctimonious Opus models.
They did a paper on this a long time ago[1], I'm sure they have altered their techniques some, but I this was the earlier signal that they really treated the AI as a separate entity and trained it as such. And focus not just on it's capability, but it's *behavior*, when to push back on the user, etc... Imo. it's more about the philosophy of the constitution in this paper than the specific RLAIHF, I'm sure this idea of "when to follow instructions and when to trust yourself is in their system prompts for all their RL environments, and even their rewards.
This idea of an objective function vs what the model should do during human interaction is a huge question and a very subtle one to get right, but it's why they were winning in real world application even before they were winning in raw benchmarks.
Especially in the era of RL environments dominating more and more compute it's vey easy to overfit a model to be very good at accomplishing a narrow task exactly as specified. Much more complicated to have it respond appropriately on turn 5 of a nuanced conversation with a user about their private codebase.
Secondly, this strategy leads to Claude models also being the slowest models - so each response / tweak iteration wastes more time than other models.
Thirdly, because Claude's intelligence drops off so rapidly with thinking levels, you end up paying way too much for marginally higher intelligence (which may not even be perceptible) for its Max levels whereas the equivalent levels of its competitors cost less due to fewer output tokens being used (and generally saner pricing).
It makes sense too, because of course Google should have a ton of exotic training data. It's probably their comparative advantage, over getting super-clever with reinforcement learning.
Yes, it's called the "Soul Document". A long trade secret document depicting everything about Claude (behavior, values, character, boundaries etc.) in great detail. They use that in training to shape Claude's identity.
I think this makes all the difference because I feel like talking a very good coworker. It understands me, helps me and slightly pushes me to be a better version of myself. This is in contrast with a robot blindly doing what I want.
Only evidence I can point to is that over a year ago I built a custom ai harness for a corp with no usage caps on anthropic models. Claude taught me all the buzz words as I built long before they were publicly discussed anywhere. From that period try searching for "Orchestrator" hits on google/github etc. My code has it for the agent with the subagent calling toolset, Claude's recommendation. It performed long horizon tasks really well too. Almost exactly like Fable, same persistence, with worse performance of course/accuracy/hallucination. Just had a watchdog script and a validation agent that ran when the main orch claimed it was done and yell at it about what it didn't complete, all on 250k context sonnet.
Now, bundle that basic format with a nice conversational LLM that Anthropic keeps pretending is its own model doing everything by itself with its much bigger context and you get what we have today.
Essentially providers specialize their AI stack + LLM personality + backend fuckery they pretend isn't 90% of model capability into their api spigot. Openai has been focused on 'better chatbot' for the most part and Google appears to have been sitting back and waiting to see what worked first.
Then, quickly with the big perf upgrade with the 'orchestration layer' path, Google had a target to quickly RL on. Thats why you should only be using Agy on Carbon with /Swarm on IMO. Its great. Carbon is fine tuned for orch work. Same reason Fable often performs worse at benchmark tasks than Opus. Its meant to be an Orch not a worker.
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?
Good lord, this is an insane flow.
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.
Said another way, I am using "Google AI Pro" free for students, which seems to be a deceptively different plan internally than "Google AI Pro" for consumers, even though they list the same features on their marketing websites.
This is an underrated quality of Gemini models.
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".
Showing 400 of 765. Read the rest on Hacker News
Comments are loaded live from Hacker News and are not stored by Mid or Real.