They have an insane amount of resources but are so bloated, have a ton of bureaucracy, and have lost a lot of their top talent. Looks like they have figured things out now though.
They have top of the line research, I mean, they invented the thing, and they are still doing stuff that goes beyond making a bigger LLM, as it seems to be the only thing the others are doing, research-wise.
They also have hardware capabilities no one else has, they invented the TPUs, and have lots of them.
They also have the brand recognition, ecosystem, user base and lots of money.
So why are they lagging behind on the actual products, and by lagging behind? I mean they are always in the position where they are just behind the frontier, but still in the game.
It seems almost deliberate. Maybe they are playing the long game, waiting for the bubble to burst. Maybe they don't want to compete against the more profitable parts of the company. Or maybe their researchers are not interested in making an actual product unless being forced to.
Spare me the celebration. Artificial Analysis should not be granting any company the favor of a #1 ranking for a model that isn't released to the public. Also, congrats to AA for handing over their private benchmark set to Google too lol cmon now.
Google's model release is pure PR puffery. It's no coincidence that it was released on the day of its earnings call.
It kind of doesn’t make sense. Modern google is typically kind of bad at building products and very good at getting them in customer hands. This is the opposite of that.
It does make sense, when you consider that Google had its earnings report yesterday, and missed its targets. $GOOG is already down 13%. It's an announcement purely for their investors.
I'll use this opportunity while it's still early to share my experience with Argon.
I got it via the "Pro" plan, exclusively for personal task usage. So no agentic programming.
For starters, the limits are dreadfully low. Five messages on medium is already enough to trigger a 4 hour cooldown. Secondly, the best part is the integration with Google search and Maps. Makes it seamless for trip planning. Lastly it's the only "SOTA" model in a long time that has actively hallucinated details. I don't know if it's because the Gemini harness is not asking it to verify information, but I've had some real Deepseek V3-like lies in responses.
Either way, I find it difficult to get invested in Gemini because nobody will have access to Carbon for a few months anyways.
Yeah... Given that everyone else seems to make models available either immediately or within days of announcement, idk why Google feels it's competitive or acceptable to do these 'available at some point' releases.
Capabilities are also improving so fast that this fomo doesn't really keep people excited. If I can't run it, I'll sorta just click off. And given how they handled the Argon release, I've not even bothered look at this Carbon one. Fully expected it to not be usable yet.
The other labs do the same process, but they don't announce the model until it is about to be released. For some reason Google is announcing at the start of th process.
My guess is Google knows it is (way) behind the and needs to keep drumming up interest. Although at some point it becomes counterproductive (raising "yet another not available model; is it vaporware" vibes).
Eh, maybe I'm too much of an AI safety person, but I'd rather the labs be forced to announce when they are working on large SOTA models like they are a clinical trial. That way if it's never announced they have to say something like "training didn't work out" or "it took over the server farm and we had to burn the hard disks".
Creating a false sense of scarcity is essential to get people talking about it. I get almost endless inquiries from peers now that I have access to Mythos.
I'm not sure about this. I myself am curious about this model, but not for my own use. I've already written it off as too expensive. The money won't be flowing their way regardless.
What's the betting we see another "dramatic" Anthropic release either shortly before the IPO, or at same time they release their Q3 results which were conveniently omitted from their S-1 filing.
It's late in life to learn this, but let me be the one to let you in on a little secret: people often say only some of the things they know, they often selectively tell you things while not revealing everything they know.
You are assuming that Carbon is ready for release, deployed to their datacenters, that all scaling and other issues have been worked out, and they are confident in all of the above. Not to mention additional internal constraints like alignment testing or other safety procedures their own policies and guardrails require.
If they are smart, which they clearly have some smart people if they are making the best frontier model, they would account for the above needs in their timeline, and announce the model when they can prove they have it but also before they are ready to release it. This is so they can be at the top of the frontier model lists, boost their brand for both Gemini and Google, get the bump in stock price and investor confidence, steer adoption decisions for people making decisions for what platform they might adopt, etc.
> argon was announced to coincide with performance review deadline. carbon was announced to coincide with quarterly earnings release.
Same as it ever was. Google is really completing their IBM-ification. We need to release these models now because year end reviews are about to hit and I need a win.
In my experience with antigravity-cli for coding Argon gets the task finished much faster than 3.8 Flash and with a higher quality output. The quotas on Pro are about 2 hours of usage per week for Argon. Its a really nice model to use. Only thing that is annoying is that it still struggles with tool use on the gemini website.
Same. I subscribed to Gemini pro and requested a refund literally the same day. Asked it to put a technical image on a presentation (gave it the image) and it kept generating a different image of a peaceful landscape... Yeah, very pleasant but totally unrelated.
hmm, I haven't tried Carbon yet. But I have been using Argon on and off via agent-shell - https://github.com/xenodium/agent-shell and unofficial ACP client.
I am also using it via Google pro tier and did not immediately run into quota issues. But as someone else said in previous discussion of this - Google Pro tier already includes lot of stuff (including Youtube premium etc), the AI plans are just a nice addon.
yeah thats basically the only reason i have the google AI plan
YT Premium, increased drive/photos data, & home premium (though i may have that already tbh)
Only using flash & argon for toy projects since its included though Argon availablity is very small, similar to the included claude usage avaialble in antigravity
I am having a very positive experience with Argon: since it was launched I accessed it through Ultragravity and connected via a Google Cloud account. For one month I have been trying to move the needle in SHA256 cryptoanalysis with Claude to not avail and Argon moved "two needles".
Since then I tried it for experimenting on diverse fields. For example, this week there was a discussion here in HN about the non-continuation of Python funding on the JIT, so I "played" again to removing the GIL and moving to the JIT in Python, it did it. I played with many other things, even the Riemann Hypothesis (it just removed their rational parameter rounding only moves 0.875 towards 0.87495702, nothing significant though).
Many other relatively great things were done this week in parallel. I have some budget for research.
Note that Argon was rolled out for IDEs for people on paid AI Pro on Oct 15/Oct 16 (a couple days after Ultra users), but of course, usage limits are extremely low
> I find it difficult to get invested in Gemini because nobody will have access to Carbon for a few months anyways.
I feel like you're being overly dramatic here. Argon was 13 - 15 days.
Argon was announced September 30, made available on October 13 to Ultra everywhere, and Pro in the Gemini App. Pro plans got Argon on October 15 in AGY.
This is the first time they've slow-rolled a model. Who knows how long they'll wait on Carbon.
The Ultra-lite plans ($99/month) get some pretty good quota where you can actually use Argon some more. Maybe Google shouldn't be grouping together plans, but it should also be noted that $20/month also gets you 5TB of storage on Drive. If you paid for Dropbox, you get 3TB for $18/month (without any AI model usage).
I have 160G of screen captures from Tumblr from my then angsty teenage kid. I think my spouse takes a few Gig of photos for each outing we go on. Several people dress for fancy thing via a process involving many clothes changes and even more selfies.
If we go back only 45 days, Opus 5.5 was not even out yet. So for more than a third of the current SOTA model's existence, Google's response to it was under its own embargo. When it was finally out, paying users get 3-4 prompts a day, which means Argon will be experienced by a tiny minority of people, another unforced error on Google's side.
By the time Carbon comes out, I fully expect Anthropic, ChatGPT and/or a Chinese lab to have its answer ready to ship in less time than it took Google to approve a release.
It’s hard to get interested in Google products because they throw them out all the time. If it’s not a step change in functionality why even go near their baggage.
You also get YouTube premium and home premium on the Google ultra plan. I didn’t realize how nice it was not to see ads on YouTube, and the home summary of who walked next to my home each day is bizarre. The grouping doesn’t make sense at all, since I would have bought it anyways, but it’s a nice bonus.
It's annoying that there is no way to just pay the marginal difference between Premium Lite and Premium for AI Pro subscribers. I have a bunch of interacting subsidies from Google Fi, Pixel phone purchases, and Google One AI Pro, but it's hard to keep them straight.
I similarly had Argon respond with a vast hallucination. It surprised me because it wasn't a little wrong on the details but a fully false verbose response that completely fabricated fabricated everything. I'd recently had Claude Code help me create transcripts for some audio notes from my mother in law and decided to see how some other models would handle the same task sans harness. Argon made up a long conversation from a police sting operation with under cover operatives seemingly just because it read the file names and identified that they would have come from a recording device.
Honestly the fiction writing was pretty impressive!
Suddenly very curious what the llm roleplay people think about this one lol. Fortunately I've already learned to stifle my curiosity on this specific subject.
It goes way back, I work in AI/Ml (taken over by transformers/LLMs) space since early days of this "AI Summer". We get some nice grad school students interning from top universities. I met one during a trip to the office, from top notch school, introduced me to Triton programming when it first came out. It was early days of LLaMA just coming, when The Bloke was regularly dropping finetunes, I was joking asking him what the hacker stack was for "AI waifus" that the cool kids were using. Lo and behold he drops the name of the SillyTavern project. Which later was really educational in setting up local inference servers and in other ways, may have had to, ahem, delete some browser history.
But my experience is that the agent matters a bit more than the model. It is fun on ChatGPT because of its recent (?) bias to always be creating images for conversations around fiction.
what are some of the vagaries of how that use case depends on the agent aka harness? i’ve only been focused on sort of how to make models reliable for my math / cs / engineering and arts and crafts work loads, but RP is an important category for some folks. like are the fails of the various models different from more technical yet creative work loads?
eg im working on rigorizing a proof that a fun algorithm of mine is not only the first provably efficient algorithm for a particular problem, but its also the fastest in practice. ive found that for the sort of stuff i need to assume the model will get confused in the out of training distribution stuff from my personal library of ideas, so they will at best be a confused secretary intern mashup
If you uploaded the audio via the web site or app and got terrible results I'd bet Argon wasn't the model doing the transcribing. The only way to know for sure is to try again with the API.
Oh it definitely didn't transcribe anything at all! This was through the WebUI. I'm certain it would have done better through Antigravity or such. Really I was just expecting a denial but I figured what the hell, the files are small and I like to see how the results of various models compare.
Yes! I was having Argon summarize some project updates and it... just made up entire paragraphs that had no basis in reality. It felt like I was back in GPT3 days.
Will have to give Carbon a try, but if it is anything like what I saw before it was entirely unusable.
That's the problem for me. Google can have the world's best model; it doesn't matter to me, because I effectively can't use it.
Both Opus 5.5 and Sol 6.1 are extremely good and provide a lot of inference on their $200 plans, and then there are the Chinese models - DeepSeek 4.1 Flash, MiMo, and StepFun all offer a lot of value per cost.
At the moment, I don't see a niche where it makes sense to use Google's models, regardless of how good they are.
In my experience, (using 3.8 flash) the Google models do a fairly good job at summary notes and simple documentation (do this, do that). It doesn't seem to be as bad as Claude for producing headache-inducing text.
Good to know. I figured I'd start with Gemini because if I used Claude, I'd probably get addicted to the results. Then I would have to commission a velvet painting of me dressed in Dickensian beggar clothes with big eyes and a quivering lower lip, holding my hands out, asking for tokens.
Google has a $99 plan that I’m using right now and I found flash to be reasonable for dev, but I spent my now past career at Google working around Gemini Flash limitations (when it was much worse than it is now) and leveraging subagents in antigravity. I want to dip my toes into anthropic or openAI models, but getting setup seems to require either another plan or expensive tokens, and I’ve found DeepSeek to be a reasonable back up/second model that I’m not eager to just go in and see what I can do over there.
Agreed on the limits, they're quite frustrating. I've still been preferring Argon to Opus 5.5 though, because even though Argon gets details wrong, I find that it's much more creative and willing to explore ideas. In my experience Opus tends to be rather condescending and won't make room for less orthodox ideas to develop. For what I'm doing, I'd much rather have good ideas that I implement mostly by hand rather than regressing to the mean and have the LLM write the code.
I have experienced all of this too. I also find Argon still seems to have the sycophancy problem that some of the other labs have stamped out to a large degree.
> Lastly it's the only "SOTA" model in a long time that has actively hallucinated details.
Yesterday I used Google AI-Search (so Gemini Flash 3.5?) for opinions on a traffic accident and it continously constructed law-interpretations and then referenced the wrong paragraphs. It talked about rules on road-crossings and cited paragraphs pertaining to alcohol intoxication as sources.
Gemini models always suffered from this. They were almost always best in terms of intelligence, but one of the worst to use in practice, because they have a lot of failure modes on simple tasks that simply give factually incorrect answers, or completely break and output gibberish.
It is expensive as hell. That said, on my weird personal benchmark (https://botbowlbench.com/) it reliably beats Opus 5.5, which surprised me a fair bit.
Trip planning is the most often used use-case for agents/bots I see in marketing. I'm lucky to travel once a year, but if I did have a trip, I wouldn't feel comfortable letting an LLM take the wheel. Too many variables for wrong flight or room, closed locations in itinerary, etc.
Now, critical server infrastructure? Sure, have at it Claude, make no mistakes!
I would literally never let an AI plan my trip. I do have very specific interests though and it's nice to have an overview of suggestions of places to go.
Plus, it's great at researching the history of places you're visiting, so you can find cultural experiences a bit easier.
I don't travel often, but I do live in Europe where it's easy to visit a new city or country in the weekend. This makes this functionality a quite common use for me.
It's terrible. I tried it via API in OpenCode. It's so slow, and the lack of reasoning output makes it feel even slower. And I also noticed the lies. I gave it a high-res image of a map to classify labelled locations, and it started going on about how smart it was for the map creator to use squares as the symbol for one of the categories of labelled places, and gave me a list of all the places with squares along with their category, but the map has no squares. It's circles, triangles, and stars.
Yeah that "Pro" usage is a joke! I wish google paid more attention to what the people want! it's a shame I loved Gemini 2.5 and 3 when they first came out they were super amazing!
>Lastly it's the only "SOTA" model in a long time that has actively hallucinated details. I don't know if it's because the Gemini harness is not asking it to verify information, but I've had some real Deepseek V3-like lies in responses.
I don't know what it is about Google models, but they are the worst at hallucinations. I don't understand how they can so consistently be this bad.
It's kind of an odd release. Yes, it's technically the top intelligence model by like one point above Claude according to Artificial Analysis, but it's like barely cheaper?
It looks like the best sort of bang for your buck is still GPT Sol and GPT Astra. Claude and Gemini have just been going for that number one intelligence spot by sacrificing a lot of efficiency.
It does sort of show their hands a bit that their larger goals might be outside of just consumer sales but also seeking hyper intelligence that could bring value in new ways of discovery (if it ends up being a legit outlet for such things consistently)
> because nobody will have access to Carbon for a few months anyways
From Google saying that Argon would be 'available soon' (as they did just now for Carbon), it took them 13 days to release it. So... unclear that Carbon won't see the same trajectory.
> For starters, the limits are dreadfully low. Five messages on medium is already enough to trigger a 4 hour cooldown.
Is that chat messages that involve tool calls (web search, file reads, etc) or actually just messages?
When I'm on the $20/month OpenAI plan, I can get about 3-5 turns out of Astra before the 5 hour limit is up.
When Fable was on the $20/month Anthropic plan, I remember getting about the same.
3-4 turns out of the $20/month plan would seem about par with other providers for their best model, but really Carbon should occupy that slot, not Argon.
If Argon's limits are that low, it wouldn't make sense to use that plan for anything when models of similar intelligence are available with much higher limits on other $20 plans.
I had been on the pro plan for a few months and recently upgraded to the $200.00 AI Ultra plan because I've gotten good mileage out of 3.8 flash while keeping in mind that it's far from perfect. Argon is better in my experience in terms of reliability (especially in not claiming to have done things it clearly didn't) but I have to agree with the hallucination part, for a "SOTA" model (before Carbon came out at least), it hallucinates significantly more even on moderately complex tasks compared to Astra or Opus while being much slower than 3.8 flash (as expected of course). Another irritating thing is that like the 3.6+ models (but thankfully less often), it has a nasty habit of sometimes recognizing unstaged changes in the branch as not being its own work (whether done manually or with another agent) and deciding to undo them or even delete files. That's something I don't expect any SOTA model to do without at least asking the user.
Gemini so far has also had terrible computer-use capability or tool-use capability compared to ChatGPT (haven't used enough of Claude for these tasks lately to really compare). I tell ChatGPT to make some formatting changes to a LaTeX file I gave it and generate a PDF, it get's the PDF generated most of the time; occasionally the download link doesn't work, but that seems to be a server or link-generation issue, the actual PDF is generated and it quickly fixes the broken link when told. With Gemini, it's a long struggle and I use LLMs for formatting LaTeX a lot, so sometimes if ChatGPT is doing some long-running task for me, I have to begrudgingly give the document formatting task to Gemini, and it's dreadful every single time.
Before Argon was a thing, it once generated a PDF file, then started claiming PDF generation was not part of its functionality and the link that was previously working stopped doing so. I think I had inadvertently jail-broken a functionality because it refused to generate PDFs again and said the first time was a mistake. With Argon, it doesn't deny having the capability, but it takes much longer than ChatGPT, and messes up more formatting than it fixes. Its PDF viewer is also trash and seems like it gets built from scratch every time you want to view a PDF.
A secret model I trained last afternoon is the actual #1 on benchmarks.
No, you can't use it. You can't even see it, and I won't provide you any information. But you must trust me that I am the genius who made the best model in the whole world.
I tried Argon for a large historical HTR transcription project recently, since I heard it was really good at this. It was really good, but like all top US models I've used it was
* Definitively very good at domain specific knowledge. It can read sloppy 19th century danish gothic handwriting with faded ink just fine. Not perfectly, but probably better than I could have done with a day of struggle for every page.
* Expensive.
* Bad at actually doing what I said. I had told it in the prompt to be careful not to miss text, but I hadn't explicitly told it that some images may have text in double-column layout. So on those images - over half of my 300-page document - it produced an excellent transcription of the first column, and then considered itself done. Want to try again with a more careful prompt? That'll be another $20.
I don't know how useful this reply will be since I haven't tried Argon specifically, but with other models I've found that the best strategy for work that is shaped this way is to have the model process it in chunks and iteratively verify and build up the rules for processing. So you have it process 5 to 10 pages at a time and self-verify it's output before continuing. Every lesson gets propagated forward to the next chunk.
To get the best performance you really need to present your problem as an iterative verification loop.
This is the only use case I've found it superior to ChatGPT or Claude for - helping me understand higher level math. It's slightly easier to understand than the other models
Gemini is the best at convincing you that you understand scientific concepts. Unless you have applied that knowledge in other contexts and been given feedback that your understanding is correct you have no idea whether it is good at actually explaining them.
That's generally true for all LLMs, but as with any tool - you can actually use it in a way that produces actual, verifiable results. Obviously it requires a lot of work that 99% of people won't put into what is essentially seen as a 'magic black box', but I do agree with GP that Gemini is quite useful in actually breaking things down in an understandable way, compared to most other models.
My experience with it is mostly using it against an actual, academic source, where I can easily tell if it starts making things up. I wouldn't rely on it for anything that I can't confirm myself.
I had a research question in my area of expertise. Tried it with ChatGPT, Claude, and Gemini. I think all three could have solved the question, but ChatGPT was an unreadable mess with too many sections and headings, Claude got stuck thinking and never produced an output. Gemini gave succinct explanation that I think was similar to ChatGPT conceptually but that I was able to follow and verify. The answer was actually fairly trivial to understand; I just hadn't gone at the problem from that perspective.
I know it is common for people to try AI in domains they aren't experts in, but it is also common for people to use AI in domains they are experts in.
This is also what keeps me using it. And whether this is rational or not, I feel that if it has a better / more concise understanding of core science/math, it also has a better shot at writing concise and correct code for those topics (a lot of what I do).
I personally find it much better than Claude in day to day tasks. Easier to understand and to the point without getting overly verbose like Claude always does. I solved some pretty difficult engineering problems with it, and most of the time I run Claude in parallel, and more often than not I found Claude to be less precise actually.
It's unobtanium so nobody should care about it. Any frontier AI company can nowadays create models with ridiculous resource and token usage that will get extremely high scores, but they're of no use when you can't use them.
I take this opportunity to give feedback on Argon. The limitations are a problem. I have a Pro subscription and Argon already stops working after 3-4 prompts with long delays afterwards. It's virtually useless right now even when you pay monthly for it. It is better for some tasks than 3.1 Pro, though. Not much in my experience but it can sometimes fix errors that 3.1 has problems with. I never use Flash models since they often give abysmal results.
I don’t care and in my opinion AA does a disservice to their users by allowing such blatant manipulation. What would happen if every AI company were to include their unreleased models like this? Put the unreleased models on a separate page and leave the headline benchmark rankings for models that you can actually use.
"We're taking a deliberate approach to rollout and are actively engaged in the U.S. government’s voluntary process for pre-release model testing. Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program. We’re continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible."
Gemini 2.5 flash and 3 flash were highly underrated models that performed way better than the peers and we used the hell out of them.
Since then none of the Google models have competed well with the others at their price points. Long context performance drop, instruction following issues, unnecessarily complex and sloppish outputs.
Argon is no different, it feels just an iteration better than 3.8 flash. I don't trust the AA ranking yet.
Even with gemini 2.5 or so i remember the models used to just simply hang in production and randomly used to collate tokes and give in the output stream vs giving one by one which was weird.
Also the guardrails were implemented in a very bad manner. By default the models used to raise a default guardrail error on basic queries.
I agree. Flash 3.8 is not good and actually quite expensive, given its tendency to overthink everything. We've been transitioning lots of workloads that were formerly on Flash to Haiku 5.5. Great model.
It has already become irrelevant to anyone paying attention. Raw scores at published benchmarks mean next to nothing in modern use. They are either gamed to hell or unsolved to the point where differences can be explained by randomness. Speed, cost per token, tokens per task and hallucination levels are the useful axes you can compare along these days if you're not doing scientific frontier level research.
They probably prefer being used by the deep-pocketed organizations that have access to Gemini 4 Carbon even if it means losing the entire rest of the market.
I was promoting Google models (more like Google ecosystem) for what Google was doing with their models.
Making it accessible to non-technical people (which is what I am doing).
e.g. NotebookLLM was one that got me interested in Google's ecosystem. It was easy to use and explain even to my parents.
Now, their models (and NoteBookLLM merging to Gemini, making it extremely hard for non-tech people and other stuff) release very slowly, their models lie their ass off, very politically skewed, etc.
Just so everyone is aware of the "coincidences" that happen with this index:
- Two days ago AA announced their v5 Index which added a few new benchmarks and tweaked a few things.
- Today the newest model announced from a big lab tops the charts.
I'm not saying that AA is definitely taking money from the big labs and changing the benchmark to make the labs' new models look better, but it kind of looks that way. This pattern of behavior has happened before, where the index gets tweaked right around when a new model comes out, and the new model is on top. AA does not, AFAIK, provide a way to see how the new models would have scored under the older index versions. You could probably calculate it yourself, but I'm not going to do that much legwork just for an HN comment.
It's interesting because they publish the index methodology, so you can't really accuse them of not being transparent, but I don't really believe that the choices to update the index were completely independent of the model they happened to be evaluating at the time.
Related: The DXOMark camera scoring method also had similar behavior, where they would release new testing methodology right before flagship smartphones were released. Again, they are technically being transparent about the process, but it kind of smells like the "independent" testers may not be as independent as the average user would hope.
Yeah, it is very suspicious. Each time they update their ranking algorithm, the closed models from US labs jump up and any previously high ranking open model jumps down. It happened with MiMo, Kimi, GLM. All were right there amongst the very top, and each time that happened, AA followed quickly with an index update that shunted them dramatically lower.
Check out vals.ai for an alternative. Not sure how independent they really are either, but it's at least good for a cross reference.
This could be another point informing your prior as to how much benchmaxxing and distilling of closed models is being performed by the open weights publishers as well.
We’re literally talking about the same thing, but ascribing different motivations.
I do think the test a new model and realize their index is broken- the new model isn’t on top, so they change the weights to make sure it’s on top.
The difference is that you seem to think they are doing so out of some innocent feeling that the new model is actually better, while I think they are doing so because they were paid to do so. I think my interpretation is more likely, but that’s just me.
I had a look at the changes and the biggest impact seems to be from adding Terminal-Bench-science 0.1 and having it contribute 15% to the overall score.
I have no doubt that these are very relevant task for people in those respective fields, but I suspect they have next to no relevance to some 99.9% of model users. Hence why especially smaller labs with smaller models didn't really train for such use cases and you can see it.
Personally, it's not something that matters at all to me, as I don't work in cheminformatics or bioinformatics.
This is also arguably more of a test of a model's knowledge base than reasoning ability.
But most importantly, there is also the question to which extent the aggregate score then remains a meaningful result when you mix up too many too disparate benchmarks.
The coding agents subindex, in contrast, looks at least superficially plausible [1].
But coding is where something like 80%-90% of their revenue comes from; few other use cases are as token-hungry, and none of those are nearly as widespread. Many users will of course use the models just as chatbots, but they don't generate much in terms of revenue. And that is ultimately what determines where labs focus their resources.
My paranoid impression was more that the updating and addition of tests — obviously essential in this game — was delayed to provide advertisement for Kimi K3.1. The same happened for K3 — and this is why it was famous — but its position had collapsed by the time of 3.1
I don't understand this conspiracy/suspicion vibe around AI benchmarks -- the best benchmarks get maxed quickly in today's world, especially if the datasets are open. We're in a situation where the AA's of the world show 0-1 point difference between models, so by definition they are NOT good discriminators. These tests NEED improvement.
Do they work with frontier labs? Probably. They should work with them at least a little so that we can get an independent read on the model quality. Do they inject new tests that specifically favor one particular lab? Clearly no, as different labs have moved up over time. Do they inject new tests that help to provide some separation on scoring? God, I hope so.
Nota bene, another change they announced was dramatically increasing the amount of held back test data; this would be a simple explanation for seeing open weights models, all of which are known to distill from frontier labs and benchmarks, drop in the rankings.
Carbon scoring below Opus 5.5 on coding highlights how magnificent Opus 5.5 is as a coding agent. With just a $20/month sub, it performs a ton of work at a very high quality.
The way benchmarks work is very very different from how you work with an agent.
The benchmarks often have ambiguous instructions, and there is no interaction with the agent. So a docker image is started, the agent is called in non-interactive mode with a prompt saying to read the instructions and finish the task. The instructions are often somewhat ambiguous and the agent has no way to ask for clarification. If it picks the wrong option out of many seemingly equally likely options that test will be mark as failed.
This is just a hunch since I have not used Argon very much, and I have not used it for coding (deepseek 4.1 flash 1017 with dsh, and qwen 3.8 max are so good for coding tasks that local models have difficulty with).
My hunch is that Google will end up being best for non coding tasks. Google can concede coding to other vendors who pay Google for infra.
190 comments
[ 3.4 ms ] story [ 67.9 ms ] threadThey have top of the line research, I mean, they invented the thing, and they are still doing stuff that goes beyond making a bigger LLM, as it seems to be the only thing the others are doing, research-wise.
They also have hardware capabilities no one else has, they invented the TPUs, and have lots of them.
They also have the brand recognition, ecosystem, user base and lots of money.
So why are they lagging behind on the actual products, and by lagging behind? I mean they are always in the position where they are just behind the frontier, but still in the game.
It seems almost deliberate. Maybe they are playing the long game, waiting for the bubble to burst. Maybe they don't want to compete against the more profitable parts of the company. Or maybe their researchers are not interested in making an actual product unless being forced to.
Google's model release is pure PR puffery. It's no coincidence that it was released on the day of its earnings call.
Argon did not leave a mark or change anything
Why would anyone care about carbon as well
I got it via the "Pro" plan, exclusively for personal task usage. So no agentic programming.
For starters, the limits are dreadfully low. Five messages on medium is already enough to trigger a 4 hour cooldown. Secondly, the best part is the integration with Google search and Maps. Makes it seamless for trip planning. Lastly it's the only "SOTA" model in a long time that has actively hallucinated details. I don't know if it's because the Gemini harness is not asking it to verify information, but I've had some real Deepseek V3-like lies in responses.
Either way, I find it difficult to get invested in Gemini because nobody will have access to Carbon for a few months anyways.
Capabilities are also improving so fast that this fomo doesn't really keep people excited. If I can't run it, I'll sorta just click off. And given how they handled the Argon release, I've not even bothered look at this Carbon one. Fully expected it to not be usable yet.
The target is to make a celebratory announcement.
This is especially true around earnings, not necessarily for some sinister reason, other than it's material information that can influence the market.
You are assuming that Carbon is ready for release, deployed to their datacenters, that all scaling and other issues have been worked out, and they are confident in all of the above. Not to mention additional internal constraints like alignment testing or other safety procedures their own policies and guardrails require.
If they are smart, which they clearly have some smart people if they are making the best frontier model, they would account for the above needs in their timeline, and announce the model when they can prove they have it but also before they are ready to release it. This is so they can be at the top of the frontier model lists, boost their brand for both Gemini and Google, get the bump in stock price and investor confidence, steer adoption decisions for people making decisions for what platform they might adopt, etc.
they don't really care how end-users will receive these models, they're just announcing for internal targets.
Same as it ever was. Google is really completing their IBM-ification. We need to release these models now because year end reviews are about to hit and I need a win.
In my limited usage, it has fixed issues which has not been correctly fixed by sol-6.1 (medium) - https://github.com/auto-gnufied/markdown-modern/commit/a189a...
I am also using it via Google pro tier and did not immediately run into quota issues. But as someone else said in previous discussion of this - Google Pro tier already includes lot of stuff (including Youtube premium etc), the AI plans are just a nice addon.
Since then I tried it for experimenting on diverse fields. For example, this week there was a discussion here in HN about the non-continuation of Python funding on the JIT, so I "played" again to removing the GIL and moving to the JIT in Python, it did it. I played with many other things, even the Riemann Hypothesis (it just removed their rational parameter rounding only moves 0.875 towards 0.87495702, nothing significant though).
Many other relatively great things were done this week in parallel. I have some budget for research.
Could you expand on this? I’ve never seen a frontier model that doesn’t hallucinate details fairly often, what do you mean by “actively” here?
I feel like you're being overly dramatic here. Argon was 13 - 15 days.
Argon was announced September 30, made available on October 13 to Ultra everywhere, and Pro in the Gemini App. Pro plans got Argon on October 15 in AGY.
This is the first time they've slow-rolled a model. Who knows how long they'll wait on Carbon.
The Ultra-lite plans ($99/month) get some pretty good quota where you can actually use Argon some more. Maybe Google shouldn't be grouping together plans, but it should also be noted that $20/month also gets you 5TB of storage on Drive. If you paid for Dropbox, you get 3TB for $18/month (without any AI model usage).
(googler, opinions are my own)
my whole family of 5 people is using in total 160 GB. personally I use about 15 GB
By the time Carbon comes out, I fully expect Anthropic, ChatGPT and/or a Chinese lab to have its answer ready to ship in less time than it took Google to approve a release.
Honestly the fiction writing was pretty impressive!
But my experience is that the agent matters a bit more than the model. It is fun on ChatGPT because of its recent (?) bias to always be creating images for conversations around fiction.
eg im working on rigorizing a proof that a fun algorithm of mine is not only the first provably efficient algorithm for a particular problem, but its also the fastest in practice. ive found that for the sort of stuff i need to assume the model will get confused in the out of training distribution stuff from my personal library of ideas, so they will at best be a confused secretary intern mashup
Will have to give Carbon a try, but if it is anything like what I saw before it was entirely unusable.
That's the problem for me. Google can have the world's best model; it doesn't matter to me, because I effectively can't use it.
Both Opus 5.5 and Sol 6.1 are extremely good and provide a lot of inference on their $200 plans, and then there are the Chinese models - DeepSeek 4.1 Flash, MiMo, and StepFun all offer a lot of value per cost.
At the moment, I don't see a niche where it makes sense to use Google's models, regardless of how good they are.
for coding .. it works ok but i prefer the openai or anthropic models, much better results.
Yesterday I used Google AI-Search (so Gemini Flash 3.5?) for opinions on a traffic accident and it continously constructed law-interpretations and then referenced the wrong paragraphs. It talked about rules on road-crossings and cited paragraphs pertaining to alcohol intoxication as sources.
Gemini models always suffered from this. They were almost always best in terms of intelligence, but one of the worst to use in practice, because they have a lot of failure modes on simple tasks that simply give factually incorrect answers, or completely break and output gibberish.
Trip planning is the most often used use-case for agents/bots I see in marketing. I'm lucky to travel once a year, but if I did have a trip, I wouldn't feel comfortable letting an LLM take the wheel. Too many variables for wrong flight or room, closed locations in itinerary, etc.
Now, critical server infrastructure? Sure, have at it Claude, make no mistakes!
Plus, it's great at researching the history of places you're visiting, so you can find cultural experiences a bit easier.
I don't travel often, but I do live in Europe where it's easy to visit a new city or country in the weekend. This makes this functionality a quite common use for me.
I don't know what it is about Google models, but they are the worst at hallucinations. I don't understand how they can so consistently be this bad.
It looks like the best sort of bang for your buck is still GPT Sol and GPT Astra. Claude and Gemini have just been going for that number one intelligence spot by sacrificing a lot of efficiency.
It does sort of show their hands a bit that their larger goals might be outside of just consumer sales but also seeking hyper intelligence that could bring value in new ways of discovery (if it ends up being a legit outlet for such things consistently)
From Google saying that Argon would be 'available soon' (as they did just now for Carbon), it took them 13 days to release it. So... unclear that Carbon won't see the same trajectory.
https://www.youtube.com/watch?v=NO2I3ueVQrk&lc=UgxRXtipA3QbA...
tldr; it's not as good as opus 5.5
Is that chat messages that involve tool calls (web search, file reads, etc) or actually just messages?
When I'm on the $20/month OpenAI plan, I can get about 3-5 turns out of Astra before the 5 hour limit is up.
When Fable was on the $20/month Anthropic plan, I remember getting about the same.
3-4 turns out of the $20/month plan would seem about par with other providers for their best model, but really Carbon should occupy that slot, not Argon.
If Argon's limits are that low, it wouldn't make sense to use that plan for anything when models of similar intelligence are available with much higher limits on other $20 plans.
Gemini so far has also had terrible computer-use capability or tool-use capability compared to ChatGPT (haven't used enough of Claude for these tasks lately to really compare). I tell ChatGPT to make some formatting changes to a LaTeX file I gave it and generate a PDF, it get's the PDF generated most of the time; occasionally the download link doesn't work, but that seems to be a server or link-generation issue, the actual PDF is generated and it quickly fixes the broken link when told. With Gemini, it's a long struggle and I use LLMs for formatting LaTeX a lot, so sometimes if ChatGPT is doing some long-running task for me, I have to begrudgingly give the document formatting task to Gemini, and it's dreadful every single time.
Before Argon was a thing, it once generated a PDF file, then started claiming PDF generation was not part of its functionality and the link that was previously working stopped doing so. I think I had inadvertently jail-broken a functionality because it refused to generate PDFs again and said the first time was a mistake. With Argon, it doesn't deny having the capability, but it takes much longer than ChatGPT, and messes up more formatting than it fixes. Its PDF viewer is also trash and seems like it gets built from scratch every time you want to view a PDF.
No, you can't use it. You can't even see it, and I won't provide you any information. But you must trust me that I am the genius who made the best model in the whole world.
* Definitively very good at domain specific knowledge. It can read sloppy 19th century danish gothic handwriting with faded ink just fine. Not perfectly, but probably better than I could have done with a day of struggle for every page.
* Expensive.
* Bad at actually doing what I said. I had told it in the prompt to be careful not to miss text, but I hadn't explicitly told it that some images may have text in double-column layout. So on those images - over half of my 300-page document - it produced an excellent transcription of the first column, and then considered itself done. Want to try again with a more careful prompt? That'll be another $20.
To get the best performance you really need to present your problem as an iterative verification loop.
Especially with the recent addition of graphic visualization widgets
It manages to strike the perfect balance between depth and style of prose that other Frontier models just can't quite seem to match
My experience with it is mostly using it against an actual, academic source, where I can easily tell if it starts making things up. I wouldn't rely on it for anything that I can't confirm myself.
I know it is common for people to try AI in domains they aren't experts in, but it is also common for people to use AI in domains they are experts in.
Comparison with other models.
I take this opportunity to give feedback on Argon. The limitations are a problem. I have a Pro subscription and Argon already stops working after 3-4 prompts with long delays afterwards. It's virtually useless right now even when you pay monthly for it. It is better for some tasks than 3.1 Pro, though. Not much in my experience but it can sometimes fix errors that 3.1 has problems with. I never use Flash models since they often give abysmal results.
So I'm stuck with using Gemini Pro 3.1 for now.
Want a private benchmark and see where you'd rank? $$$.
Want to show up day 0? $$$$.
Want to show up day -0? $$$$$.
Is this API-only or US-only for now?
"We're taking a deliberate approach to rollout and are actively engaged in the U.S. government’s voluntary process for pre-release model testing. Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program. We’re continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible."
Since then none of the Google models have competed well with the others at their price points. Long context performance drop, instruction following issues, unnecessarily complex and sloppish outputs.
Argon is no different, it feels just an iteration better than 3.8 flash. I don't trust the AA ranking yet.
Also the guardrails were implemented in a very bad manner. By default the models used to raise a default guardrail error on basic queries.
I used to defend Google's models months ago, but switched to Luna for less intelligent models.
But with this bs of not making it even public, pushing off release of newer models late, I lost faith in Google.
e.g. NotebookLLM was one that got me interested in Google's ecosystem. It was easy to use and explain even to my parents.
Now, their models (and NoteBookLLM merging to Gemini, making it extremely hard for non-tech people and other stuff) release very slowly, their models lie their ass off, very politically skewed, etc.
- Two days ago AA announced their v5 Index which added a few new benchmarks and tweaked a few things.
- Today the newest model announced from a big lab tops the charts.
I'm not saying that AA is definitely taking money from the big labs and changing the benchmark to make the labs' new models look better, but it kind of looks that way. This pattern of behavior has happened before, where the index gets tweaked right around when a new model comes out, and the new model is on top. AA does not, AFAIK, provide a way to see how the new models would have scored under the older index versions. You could probably calculate it yourself, but I'm not going to do that much legwork just for an HN comment.
It's interesting because they publish the index methodology, so you can't really accuse them of not being transparent, but I don't really believe that the choices to update the index were completely independent of the model they happened to be evaluating at the time.
Related: The DXOMark camera scoring method also had similar behavior, where they would release new testing methodology right before flagship smartphones were released. Again, they are technically being transparent about the process, but it kind of smells like the "independent" testers may not be as independent as the average user would hope.
Check out vals.ai for an alternative. Not sure how independent they really are either, but it's at least good for a cross reference.
At the previous iterations they were accused of favoring OpenAI. Now they are accused of favoring Google.
I do think the test a new model and realize their index is broken- the new model isn’t on top, so they change the weights to make sure it’s on top.
The difference is that you seem to think they are doing so out of some innocent feeling that the new model is actually better, while I think they are doing so because they were paid to do so. I think my interpretation is more likely, but that’s just me.
Example problems are:
- https://hub.harborframework.com/tasks/terminal-bench-science...
- https://hub.harborframework.com/tasks/terminal-bench-science...
I have no doubt that these are very relevant task for people in those respective fields, but I suspect they have next to no relevance to some 99.9% of model users. Hence why especially smaller labs with smaller models didn't really train for such use cases and you can see it.
Personally, it's not something that matters at all to me, as I don't work in cheminformatics or bioinformatics.
This is also arguably more of a test of a model's knowledge base than reasoning ability.
But most importantly, there is also the question to which extent the aggregate score then remains a meaningful result when you mix up too many too disparate benchmarks.
The coding agents subindex, in contrast, looks at least superficially plausible [1].
[1] https://artificialanalysis.ai/agents/coding-agents#coding-ag...
Do they work with frontier labs? Probably. They should work with them at least a little so that we can get an independent read on the model quality. Do they inject new tests that specifically favor one particular lab? Clearly no, as different labs have moved up over time. Do they inject new tests that help to provide some separation on scoring? God, I hope so.
Nota bene, another change they announced was dramatically increasing the amount of held back test data; this would be a simple explanation for seeing open weights models, all of which are known to distill from frontier labs and benchmarks, drop in the rankings.
The benchmarks often have ambiguous instructions, and there is no interaction with the agent. So a docker image is started, the agent is called in non-interactive mode with a prompt saying to read the instructions and finish the task. The instructions are often somewhat ambiguous and the agent has no way to ask for clarification. If it picks the wrong option out of many seemingly equally likely options that test will be mark as failed.
I get that you're saying the model might get a higher score because of luck, but at the same time you want the model to be "lucky".
I wish Argon were as lucky as Opus.
My hunch is that Google will end up being best for non coding tasks. Google can concede coding to other vendors who pay Google for infra.