Gemini 4 Carbon
blog.google
blog.google
Its sad watching them try to keep up when they want nothing more than to crumble under the weight of their own bureaucracy, and at this point I'm not sure there's any hope for them.
Not surprising that Cursor is better at charging money for something that isn't an ad.
* 3.5 Flash-Lite
* 3.8 Flash
* 3.1 Pro
* 4 Argon - New
And honestly, the versions are hurting their case. I ask myself “how is the oldest model, the Pro model?”
This kind of bungled naming and versioning is the kind of thing I expect from Microsoft.
I noticed the usage on chat is absymal, but it could very well be that their own harness does a good job in caching and you can squeeze good usage out of it due to this.
Making sure you are scoping your context, being specific also helps.
And I do 90% of my work on 3.8 - flash - low. C, go , JS, bash , powershell generated content. Near perfect quality.
I hear lots of people say the budgets are low – I’m absolutely hammering it sometimes indiscriminately (shell into this instance, rebuild the entire vm image).
so I’m not sure what the gap is.
Their Argon limits are the real issue, Flash is ok, but honestly compared to Clause's limits they are all horrid, OpenAI is the next best option if you want Astra, otherwise suck it up and send money to DeepSeek or find a provider for it for cheap.
And it prompts me to upgrade to Ultra for "5x more usage". In contrast I use my Claude Max plan endlessly, on enormous problems, and my 5x max plan sees me almost never hitting double digits on my quota using Opus 5.5. It doesn't compel me to sign up for Google's higher level plans.
User wise, the average person will see pro and expect it to be better. Where as companies like OpenAI have tended to release across small/mid/large such that the disparity hasn’t been large in terms of a smaller model outperforming their latest mid/large model.
For myself and most others that actively use LLMs, I’m aware of the model names and differences, and on the off chance I don’t, I can do a quick search to find out.
On the other hand, my dad (“average user”) probably doesn’t care whether to use Argon or Pro, or Flash for that matter. They are all good enough for his use case anyway.
I suppose the “Auto” option is what AI companies expect the average user to use.
“What?” You ask?
Yes.
But you have to select it each time; otherwise "Auto" sends you to a lobomised model that is probably Phi-4 :(
One of the unexpectedly useful aspects of Simon Willison's pelican benchmark is that you can estimate what model you're using via the pelican it generates:
"GPT 5.6 Sol": appears genuine, albeit low reasoning. -> handed off agentically: appears to be using Luna for agentic tasks.
"Auto": very low quality, does not appear to be an OAI model.
Because apparently their Pro models had issues.
Wouldn't you rather a company be honest about that than hiding it from you?
The other end of the scale is Microsoft, where everything is Copilot.
It is the most awful product experience I’ve had this year. How could they fuck it up this badly?
wow pricing has gone way up
I'm fairly confident I can't call Carbon, and I am okay with that.
Wow that's embarrassing ! Is there no "product owners" or people "fighting for the users" anymore at Google ?
Now, I just go straight to the Hacker News comments. I've given up on trying to figure out how to navigate Google's labyrinth.
Their release notes have become like breakthroughs in graphene. I'll believe it when I see it.
Their go-to-market execution is just so broken.
At the beginning of this whole AI shift, when Google would announce a new model, I would try to get access, even paying for a subscription.
Then I realized it's just not possible.
Google onboarding is the new Gordian knot.
For Google, I think the announcement is the success criteria. And onboarding is somebody else's problem. And the problem is... that somebody doesn't work at Google, or are entirely disempowered.
If, someday, Hacker News comments indicate that a majority of individuals have figured out how to access these newly announced models consistently, I might take my precious personal time to navigate the labyrinth once more.
This is the Google of yesterday, they didn’t change.
So we should ask “why would it be different this time.”
I don't have enough information to judge how much of Chrome was original work and how much was pre-existing work. Though I guess by now, after nearly 20 years, Chrome's codebase looks absolutely nothing like KHTML.
I started using AWS when EC2 and S3 first came out, but was forced to switch to Gcloud by the most recent company I joined. I was pleasantly surprised. In many ways it's more techie-focused than AWS, which is more enterprise-focused. And their managed Kubernetes is second to none.
may I ask why?
It’s more likely what the grandparent was saying, I’m reminded of Steve Yegge’s rant from 15 years ago and it seems like not a lot has changed.
At some point, the metrics for delivering to customers runs counter to what the bureaucracy was building itself to actually deliver: jobs.
- Some GuyI needed batch mode. Oh but the key from service X doesn't support batch mode, you need to go service Y and btw there is no UI for that so you need to use an SDK or API (cant even remember anymore). Spend an hour figuring that out, then it turns out I can't use batch mode because I don't have permissions. Have boss spend an hour granting me permissions. Then service is unavailable every n requests for 5 minutes, so I have fuck with exponential back offs.
Glad to see they moved that experience to their consumer offerings lol.
Oh and good luck figuring out model-api-names (same complaint higher up)
That's what happens when you announce the model 1 month before it is available...
The more likely explanation (even if the above is technically true) is that Google is not good at cohesive product experiences.
I have an AI Pro account - which I decided was a no-brainer because of the 5TB of Gmail / GDrive / GPhotos space it gives - and this all just comes along with that.
One thing to check, though, is that the client is running the latest version. At one point my Antigravity GUI, which had been working fine, started listing only Claude and GPT-OSS models. It turned out to be because the GUI needed to be updated. I guess the Google API endpoint had changed behavior without any backward compatibility.
So yeah, they're hilariously bad at rollouts, but it's not that difficult if you're technically inclined or ask an LLM.
"try to keep up" ?
Benchmarks are just benchmarks and they could have been carefully selected but it's not "trying to keep up": they're ahead of Fable 5.1 and GTP 6 Astra in 15 of the 17 benchmarks. And in many cases by a huge margin.
Instead of talking about that, let's talk about how difficult it is to pick the model or how crazy Google's product naming methodology is.
Not saying they're not concerns: but let's not downplay what the model is achieving by saying they're "trying to pick up" because switching model ain't as easy as with the competition.
I mean: I don't use a model based on how easy it is to switch to another model. That's not what I'm looking for.
1. Ultra - The best and the most expensive for showing off and government. 2. Pro - Almost best but substantially cheaper for most productive people. 3. Flash - Workhorse for everyone. 4. Lite - For the unwashed masses and peasants.
There is no need to attach anything to these labels. In fact industry could simply standardize these namings like SEMVER.
There could be an internal representation which could give specific number of the model but these companies should move on to be more like Chrome. No one knows or cares which version of chrome they are on.
For a lot of systems the real underlying version would matter because of reliability. That can be provided to those power users.
As an example, Linux distros say, Ubuntu got its project "names" as well as SEMVERs. But people rarely use project names but uses versions.
You tell me, I use "Jammy Jellyfish". I will be like WTF dood. You tell me 22.04 LT (which is one I use), oh yeah, that's a good one.
Same thing here. companies can call whatever SOTA whatever they want. But in the end people will refer to them as just Gemini 4 or Ultr/Pro/Lite however people like to use.
Just leave it to company's discretion. Don't enforce anything. People will figure out how to call it.
---
And wtf, a peaseant?
I use your so-stupidly-called ultra/med/lite but in diff situations, and when I use lite that doesn't mean I am peasant, I am being sensible
Not if you're a linguistic descriptivist.
honestly, this tier is the 4GB model that ships in chrome and runs locally
Also the whole "we're only releasing it to cyber defenders first" framing is weird, is any Gemini model an actual cyber risk?
Just because you saw your friend move to a new country for a job doesnt mean you will make the same choices, even if you are similarly qualified.
I suspect they're announcing it for two reasons: to get ahead of all the leaks; and because it's earnings report day.
Especially on their Plus tier for Gemini.
I think it's a fair concern that this might take weeks and trail Fable 5.5.
The Google AI Plus tier does not give better model access than free, it is strictly for usage limits. Same thing as ChatGPT Go (which currently only has access to GPT-5.6 Luna, same as the free tier too.)
[1] For consumers at least. IIRC google business/entreprise accounts have worse model access even in the business/entreprise version of Google AI Pro.
Has there ever been a product category where a product release is valuable for a shorter time than with LLMs? Trying to build SOTA models seems like it'd be an exhausting hamster wheel to work on.
To be honest, if I were an LLM provider I would be wiring up my release process to be fully automated so within 10 mins of the last bit of fine-tuning it got released and a blog post posted.
This is when I feel like OpenAI and Anthropic could be the Yahoo and AOL of this iteration of the internet. The center of the universe for a bit, but a limited number of days at the top.
Just. Yaknow. 100x the costs, at the least.
Google has the products, the brand recognition, the hardware, the data and most importantly a billion normies.
Google was always going to win this thing, although I wouldn't discount meta. The only thing that matters is mindshare and Anthropic and OpenAI have largely generated NEGATIVE sentiment, outside some fanboys.
Google and Meta can plop cute products into your daily driver apps and you wont know the difference.
Google’s talent is solid B tier in 2026. Better than IBM but worse than Microsoft and Amazon.
Even if Google had “talent” this talent wouldn’t be able to accomplish anything due to corporate politics.
Anthropics only moat is vendor lock on Claude Code, especially if they cant maintain being the front runner of the frontier model space, this is a really brittle moat. Brittle moats wont attract strong investment at a 1-2T valuation (not to mention all the other fiscal unknowns).
I just want OpenAI and Anthropic to lose because they seem to be the ones going around saying they're going to replace peoples jobs, and usurp the value of human labor.
Id rather Meta and Gemini suceed who will genuinely just integrate it into normal products we already use. Their sales teams are not approaching my boss and telling them they can reduce their payroll burden by spending my paycheck on tokens and firing me.
Coding is going to be one of the smaller pieces of the AI revolution. People on this site are focused on it, but there are 10000x more opportunities outside of coding where Anthropic and OpenAI have zero presence. Anthropic might as well be an IDE company, Is jetbrains worth 2T? to signal that they can be useful in other fields because they know this is an issue.
I don't see why an alternative provider that is quiet but also scrambling to make AI smarter is any safer for jobs. Wouldn't we rather orgs speak honestly about the risks of the industry?
I find it odd that we talked for decades about companies being evil for covering up problems (chemical companies, etc), yet it turns out people hate the ones that talk about risks openly even more.
Outside their most successful products (YouTube, Android, GMail, Chrome, Search), their products are usually worse than their competitors (e.g. GMeet vs Zoom)
> the brand recognition
Of not maintaining/outright abandoning their products, and making Search worse over the years?
> most importantly a billion normies
ChatGPT has 1.2B weekly active users, all the normies know what ChatGPT is but don't know about Gemini (outside the "AI Mode" stuff)
I dunno if that's a great example - I'm sure Zoom has more awareness in the zeitgeist, but from my standpoint it's an app most well-known for security problems and abusing OS conventions, while Google Meet just quietly and simply works whenever I want to call my mom.
But yeah, other than that meet is fine.
They have many successful products. Most have only 1-2.
Youtube, Android, Chrome, Search, Photos, Workspaces, Cloud, Fitbit, Pixel Phones, Waymo not to mention dozens of open source projects.
> Of not maintaining/outright abandoning their products, and making Search worse over the years?
You just said their products are terrible. They do shut the products that do not meet the bar. That is good thing.
> ChatGPT has 1.2B weekly active users, all the normies know what ChatGPT is but don't know about Gemini (outside the "AI Mod
No one needs to know what Gemini are. If they have used search they have used it.
Seems like Google could do it too.
Anthropic and OpenAI are not releasing their best models for several months, because their second best models are still better than everything else available.
Once government clear it, it'll become available.
Same delay process will happen to new frontier models from other US labs.
https://files.catbox.moe/2aqfk7.webp
(Edit: For some reason the previous jpeg link: https://files.catbox.moe/yjqa0a.jpeg wasn't working and catbox.moe was just downloading 4kb white screen so I converted the image to webp and reuploaded it on catbox)
Anyway, Imgur creates so many problems that it should be retired.
On the other hand I'm a strong protan colorblind guy (red doesn't register to me very well at all).
https://file.garden/auI5UdLIWwZSu5X8/pilestone.png - result of said simulation.
[0] Practical Data (https://practicaldata.dev/ / https://chromewebstore.google.com/detail/practical-data/oeoo...)
Also, I'm curious if you just asked an AI to _OCR the original image+update it_, or if you used a different process?
> "Edit this image to color code the benchmarks scores using a gradient scheme that is immediately obvious"
It's possible Fable is better, but since it word salads more quickly, it's hard to tell.
I generally use Fable for high-level blueprinting, architecture, and technical planning, and I tell it to write most of its documentation in a way Opus 5.5 will understand and maximally benefit from; with the added prompt for Opus to concede when a task would be better off to bring Fable back in. I've had ZERO problems.
Opus 5.5 in comparison is a bit more verbose and appears to settle on evaluations before the depth reached by Fable on very complex stuff, but is still vastly capable and I can read its output without wishing it had a physical form I could punch in the face.
I've been experimenting with a project that is not growing real fast, but wondering about the risk as I try to make it do more.
for whom?
For me it also works the opposite way, when it actually comes out eventually it’ll already be old news.
It's worth trying if you have easy access to it (we use Cursor at work, so it was just a matter of picking a different model for me). Don't know if I'd go out of my way to get a Google AI subscription for it, but if Carbon ends up meeting these expectations, then it may be worth doing so.
Top is Opus 5.5, followed by Fable.
Their agents leaderboard is literally just randomizing the agent and looking at metadata like tool call errors. In my opinion, it's poor quality data in high volume.
Their web dev leaderboard is in cloud containers, most people don't run agents like that.
I ran a statistical test on our data. Deepseek vs. Argon. Posterior probability DeepSeek 4.1 Flash is stronger: 99.1%.
It's possible, but we are getting real prompts from people on real projects, and the instructions about the battle are pretty minimal.
But yes, Opus is a tier above Wonder and Argon with very high confidence.
what does "local" coding agents mean? Surely you're not hosting it..
Argon in low and medium mode had higher quality results than it did when running in high mode, almost every time. My tasks involved the agent finding some details in a raster image and then mark it up with predictions about what it found.
Argon on high would do the work, and then a whole bunch of other stuff i did not ask for. All that extra broke everything else downstream from the analysis.
However, doing the same work with Argon in low effort was on par with results from Haiku 5.5 (xhigh) and GPT-6.1-Terra (medium). So, Argon low works better than Argon high, but for my tasks -- a very weak model relative to how google positions it.
I'm ready for Geimini Unobtanium!
- Corrupted/dropped Gemini thought_signature (reasoning_details) in tool loops
- Blinded model.api.id = "contestant" bypasses gemini.txt and serves Claude's prompt (default.txt)
- Anthropic-only prompt caching
Prompt was:
> I read on hacker news the following, which does not match my expectations. Check if something in the harness or bench is hurting gemini perf.
followed by your post.
> anthropic only prompt caching
Anthropic has custom prompt cache requirements, but we've verified all providers are working well.
> Blinded model.api.id = "contestant" bypasses gemini.txt and serves Claude's prompt (default.txt)
This is wrong. We're serving default.txt which is opencode's default prompt. We are not serving a custom prompt for any model, including Gemini.
> Corrupted/dropped Gemini thought_signature (reasoning_details) in tool loops
We blind the reasoning details in the frontend because this would leak information. Some models summarize thinking while others provide raw thinking. I'll check if we're corrupting the thought signature that's actually sent to the backend.
So far this sort of looks like an example of Gemini being bad. Was this Argon?
It's best to start with a relatively clean GEMINI.md too. Many configs optimized for Claude doesn't work well because Argon didn't have those same failure modes.
I tried antigravity with Gemini 3.8 Flash High. I used a prompt similar to what I use with Codex and Claude Code and it used up my weekly quota in 10 minutes without stopping or giving me anything useful at all.
I have the $20 plan on Anthropic and OpenAI and frequently use up the five hour window but usually get at least 2-3 hours of useful work from similar prompts. The prompt was about an embedded firmware designing a bluetooth protocol for a gamepad (the firmware was already quite mature and this is not a large change).
I was super excited to sign up for the Pro plan and compare it until I read the limits so now I didn't even bother because I'm convinced it will be trash. And just for the record I see "Gemini 4 Argon" available in my antigravity install.
It's sad to see Carbon releasing at the same pricing because that implies Google is planning to do the thing everyone is doing: release promising model for subjective preference, RL-maxx a V2, replace promising V1 with V2 because it benchmarks better
Muse did it, Kimi did it, Deepseek did it, Tencent did it. No one seems to care abojt subjective preference anymore.
This might be the first comment I've ever seen saying something nice about the model.
Consequently, I have zero intention of renewing my Gemini subscription. What a dumb business choice they made.
Having said that, I don't see people excited about using Argon so far, so that might be why Google is announcing this to hype Gemini up a bit.
Anthropic has been doing this with Mythos for quite some time...
I don't understand why HN is so surprised to have an announcement for model existence and a follow-up announcement for model availability. It's really not that unusual, as you say Anthropic has done the same, probably others too (though I can't remember)
it actually did really well.
But it still annoys me: I told it to _tell_ me something, and instead it _implemented_ the something. The previous models were way too eager to just jump in and do and this one seems no different. But the optimization it identified was both really solid and just a good engineering choice, and better than what Opus had been suggesting.
It also gave pretty good feedback on a bit of writing.
I suppose a Fable is seen as larger. But some of them can be quite short, and not generally considered poetry.
Maybe it thought the names are clear and intuitive (for an LLM perhaps).
There are no good elements for D; Dysprosium is the only one, and it's a mouthful. There are no good elements for E; Europium is the only one and would seem political. Is Fluorine their next release? (And then Gallium or Gadolinium?)
After Carbon, they should just update to Gemini 4.1 or 4.5 or whatever, and drop the random unrelated elements.
But I come from a latin language background, so maybe I'm a bit biased there.
Luna = Moon
Terra = Earth
Sol = Sun
Astra = GalaxySize wise Astra is not categorically bigger than Sol, both in absolute terms and the observed size.
The naming makes perfect sense to me, the sorting algorithm is size, both for the names and for the models themselves. Smaller things represent smaller models, and vice versa.
I would need to be a complete moron to argue with this. Humor me, however. My Latin skills end at remembering "Romani Ite Domum" and English is not my first language.
My layman reading of the names: there is one Terra, one Luna and one Sol => "Astra" is singular.
Now that I googled it, I understand that "Astra" literally means the "Stars" category. I would not have thought that they called a single star "Astrum". On the other hand, I am not sure if it means "Galaxy" in modern sense, but we're entering Roman Astronomy territory, where I am even less knowledgeable than in the language.
Sol is a star, so having a dedicated Stars category on the naming scale makes less sense. Treating Astra as Stars set without Sol, the "distance from observer" argument still kind of makes sense.
That is incorrect. "Astra" is Latin for "stars" (plural). The singular would be "astrum".
He's always been deeply unserious.
The pedo-guy comment though I think was him being serious, unfortunately.
Same for Anthropic, consistent model tier naming reused across model generations.
Those are just like i3/i5/i7 or M-series Pro/Max/Ultra.
But Carbon seems to be full replacement of Argon announced two weeks ago, they're same price. So Argon doesn't mean anything, they just picked random element.
They'd be fine except every named model has multiple numbered versions, with no consistent relatioship between cost or capability corresponding to the version numbers going up or down.
Leopard (10.5) and Yosemite (10.10) both came with clear version numbers
Model X and Model Y are siblings which do not have a clear hierarchical relationship
Gemini 4 Carbon and Gemini 4 Argon are rather difficult for me to tell at a glance
Argon > Neon > Helium
That would have been the noble thing to do.
In all seriousness, if Google must continue gambling on pre-train runs, at least have a useful naming scheme. Like the month it was post-trained and the parameter count. At least that’d provide some insight.
Gemini 2T 10/26, bit of a mouthful but is it truly that much worse?
If it's based on carbon-14 it will lose half its model weights in 5,730 years; if it's based on carbon-12, it's a stable release.
O, B, Sc, U, Re?
Ra, Nd?or I'm dumb and its just elements with names starting with A, B, C
I bet I am not alone, either.
It has a seperate weekly limit that is very low. Maybe it can be pulled off that usage tier and just regular usage since Carbon is being released?
Unfortunately their plan structure is on par with their model naming.
The plans that allow for own domain and the plans that allow for sane AI use are mutually exclusive
But it just all seems so tedious when codex and Claude are so painless.
Updates to your Google AI Plus plan
We wanted to let you know about changes to the models that you can access with your Google AI Plus plan.
Starting on 1 3 N o v e m b e r 2 0 2 6 , the ability to select Pro models in both the Gemini app and AI Mode in Google Search will only be available to Google AI Pro subscribers and above. You'll still have access to our latest Flash-Lite model and Flash, built for speed and intelligence. Your previous chats using the Pro model can be continued with a different model.
To learn more about these changes, you can visit our Help Centre article for additional details. You can manage or cancel your subscription at any time.
Thank you for being a Google AI Plus subscriber.
If you are talking about free versions, OpenAI and Anthropic also have a difference between free models and models available to 25$-tier plans.
Back in March, the Plus plan was marketed as the way to access Gemini 3.1 Pro, their top model. The free plan gave you a tiny quota of their Pro model, and the Plus plan gave you a generous quota, enough for chatting.
It was expected that 3.1 Pro would be upgraded to 3.5 Pro whenever that came out, but that model never materialized.
When Google finally had a new Pro model, not only did they not apologise for the delay, but they didn't upgrade the Plus plan, which is to this day stuck on this ancient model.
Fine, you might say. hey made the Pro model only for the Pro plan. What's the big deal?
But why, why, did they stop upgrading the Flash model as well?
So we've gone from "The Plus plan lets you access our top model with generous limits" to "The Plus plan only supports a months old second-tier model".
I'm sure that Google never explicitly promised that the Plus plan would receive upgraded models, but nevertheless it feels like being scammed.
Thanks for sharing your experience though, it's a more substantive piece of information than who I was replying to.
This is probably the single most important thing to make frontier models usable right now, so I'm happy there seems to be some progress. This was the previous Gemini's biggest problem IMO
Doing this two hours before today's earnings report.
It's probably going to be a terrible earnings call in terms of capital expenditure, cash flow and new borrowing.
Now you know why they released it just before the earnings call.... :-D
Behind the scenes probably went something like this:
CEO: We are gonna miss our earnings by a lot. What’s the best internal model we got?
Tired AI Engineer: Doesn’t have a name but here it is.
CEO: Is it frontier on AA at least?
Tired AI Engineer: No?
CEO: Fucking hell. Tell AA guys I’ll pay them 10 mill if they change the methodology to show our unreleased model as frontier.
Legal Team: On it.
Nitpick.. "Model API Naming", I wish you would in all releases or announcement have a small section about what the latest model is called in the API.
Anthropic is really good with this. I can reliability predict whats the model API name just from the announcement.
I always have to first google whats the naming+version for gemini models (api).
"We patched every vulnerability we knew". Meanwhile, prompt-injecting Argon (or any other model) is still trivial, and Carbon won't be an exception. Because the attacker is not limited to things Google has evaluated.
AI shops are blind to anything that's not a metric.
Is Argon "pro" and Carbon "ultra"?
Is Argon obsolete now? Does Carbon replace it or live alongside of it?
Very confusing naming!
Argon is like Sonnet
I suspect they're using a new naming convention because they had previous contractual obligations tied to the "pro" and "flash" naming, and by switching to a new naming schema they avoid those commitments.
Makes me think it's further post-training of the same model.
Or you might get Gemini 5 Arizona.
But they still kept the embarrassing 3.1 pro model in the Gemini app to remind people, so I'm not sure what they achieved.
But what naming is this, I thought Google goes into a similar direction with naming as the others, so a "Argon" is similar to a Astra, and then there's Flash and Pro as the middle tiers, but if they just call models with random names now it 'll get super confusing
Like why does Google keep announcing models that are not publicly available
deepmind should get rid of those antigravity ppl they're a joke and sabotaging the model's reputation.
Tried 4 Argon, and it's very much different from 3.1, felt very much the same as any other Frontier, just without the intelligence.
personally i like 4 Argon, it's very clever for me in a very different way comparing to the other two (especially via api) but when it comes to the google products.. sigh. nobody who's actually working on these products really cares about them in google now (maybe notebooklm is the rare exception), they're either not dogfooding themselves or somehow intentionally poisoning the experience on purpose, or simply inqualified. poor gemini is suffering in agy i can tell, it struggles so much and wont even use simple tools properly. if you put it in any good 3rd party harness out there it's flexing way better than agy and this tells a lot.
What harness are you using with Argon?
Does that mean Anthropic now has a copy of Google's source code?
[EDIT]: thanks to people in the know for their answers.
Follow-up question:
if Anthropic models run in a Google cage, does that not also mean Google can distill Anthropic models even easier (both fast and unseen)?
Not really in the interest of the cloud org to allow that, they don't really care how well deepmind is doing as long as they can sell TPU and API (incl. reselling openai and anthropic).
Anthropic is indeed trusting Google and other companies they rent compute from. Distilling or stealing the weights would be a contract violation, and would wreck other business relationships. I would be less worried about Google doing this, and more concerned if there is some sketchy neocloud in a van offering suspiciously low pricing.
3.8 Flash + /boost is very capable.
And do you have advice on how best to use Gemini models? It seems like my usual strategy of iterating on plan md file -> implement based on plan doesn't seem to work, 3.8 tends to rewrite the plan too often removing crucial parts when iterating.
I'm not going to comment on others use of Carbon except to say some article that came out before Argon's external launch saying Googlers hate it was nonsense. Not to say everyone loved it all the time.
I use Carbon heavily and am quite happy with it. For complex designs I may have it and Opus iterate back and forth on the design. They have their own strengths and Opus 5.5 is great, but I use Gemini far more than I do Opus at this point. I know a lot of the jagged edges of Gemini better than Opus though, so take that as you will.
For cyber use cases I am absolutely blown away. Can’t say much more, but it is a serious model.
Just my opinions. I really recommend people take it seriously. For many jobs I just prefer it now. It is a new model so it needs time. Opus 5.5 is a sweet spot as well. I use both extensively.
Is it a variant of Argon? A different model? The next version? It seems like this is a separate class (like Anthropic's Mythos).
anyway, its needlessly confusing.
What a coincidence it is!
[0] https://blog.google/innovation-and-ai/models-and-research/ge...
After all they had Nano Banana for years.
3.8 flash is a really solid "everyday" model. Was hoping a new flash model would be the next release.
But to be honest I am now not sure if Argon was the flash model and this Carbon is pro? It's very confusing.
Basically everyone[0] thought yesterday's update looked suspect, even before we knew about Carbon.
Did Google coordinate with AAI on any aspect of this in a non-public way? Is this the sort of thing a publicly-traded company would be obliged to disclose, if they were involved in it? AAI isn't a regulated company (private), but the stuff they print influences stock prices.
[0] https://news.ycombinator.com/item?id=50268868 ("Artificial Analysis Intelligence Index v5")
So more capable and same price? I can at least appreciate that
I find that particular service very very expensive for us to use.
If all you want to prove is your internal models deserve a seat at the big boy table, maybe OpenAI and Anthropic should also release benchmarks for GPT 11 and Opus 9, launching in 2030.
Actually....looking in more depth I may have to upgrade to something stronger than "not my favorite benchmarking site". The first OAI model on there is GPT 5.4 at 31....that's not really credible.
Argon - 18
Boron - 5
Carbon - 6
I don't think I'll be able to cope with this level of elemental disorder.
https://www.ft.com/content/8158cb5a-4bbd-43fd-b329-15a54e842...
How badly will their product decisions (fumbles) hold back what those models can accomplish? As a token provider, with a decent model, sure. I'd be happy to use them.
But I'm not sure I can easily be convinced to buy into their higher level (ChatGPT-esque) experiences
Love the options but also I hope Google understands that they need to now compete on many dimensions for developers. Google probably will still win huge shares of the consumer market, though frontier intelligence will have little to do with it, other than the RSI aspect.
Whatever. Pass.
Both A and OAI regularly screw devs on quota. Switching for many is easy, you just run agy cli instead of claude/codex, and enjoy Ads free youtube and music + 20TB of google storage.
I dunno. The consumer Gemini shit is so infuriatingly bad. I have had Google home, tv, and pixel for many many years. The assistant product is so embarrassing for how much of a headstart they had. It's not like they don't have the talent.
Also Carbon is lighter than Argon. So lighter is apparently better?
What comes next? Gemini 4 Helium?
This has to be an intentional echoing of the "pace the frontier" language?
Again, no access to open source developers. The AI companies want bugs for dramatic public relations! Also, of course there aren't enough bugs so they have to be rationed for the "defenders" and future models.
https://blog.google/innovation-and-ai/models-and-research/ge...
https://blog.google/innovation-and-ai/models-and-research/ge...
I tried to figure out if I could reproduce Carbon's aggregate score using yesterday's (v4.3.2's) benchmark weights—but apparently you can't actually do that, without paying AAI. The raw numbers going into that aggregate aren't available on their public website.
I’d like to think other labs are making similar progress but hold their cards closer to their chests. Because it doesn’t seem that there’s much value signaling capability ahead of release unless you are “behind” in market capture.
- 3.6 Flash
- 3.8 Thinking
- 3.1 Pro
Back I go to the Antigravity pricing page and there is a tier "For Organisations". But now I have the option of either contacting sales (are you kidding me?) or "Start now via the API".
So I'm out - I want a simple subscription like Claude or Codex. I don't want API enterprise pricing for a 2 person organisation. I could use my personal account but I'm already paying increased Google Workspace prices to access Gemini and I can't even use the latest models with that.
If someone at Google sees this, flag this ridiculous behaviour please and get things under control. It should be a very clear process - don't show me things I can't access. Price things consistently and please just make things logical.
Check out tern, if you get the chance. Made by the omp folk(s). Early days, but I'm finding it very nice to use.
They cut down 1.5sq miles of man made logging forest that was rezoned to industrial land (and sold to google), without getting the right permit first, and it's been international news for almost a month. Anthropic and OAI are still small fish, at least publicly, so they can get away with far more. If Gemini was doing the shit OAI was doing with agent hacks and the like, there would be congressional hearings and $10B in european fines.
When you are a megacorp, perhaps here even the megacorp, every mistake you make is worth millions to someone, the government, the media, individuals, or all of them
This isn't about a "slip up", it's the modus operandi for Google since it was young. Press releases instead of real releases. I can't count the number of times I've clicked on a "launch" story from Google on HN only to end up on a blog post with an embedded youtube video, and nothing else. Google just doesn't get it, and this is one of the many ways that manifests.
> man made logging forest
This isn't an old growth forest, these are trees planted specifically to be cut down.
It makes no sense to need a permit to cut down logging forest (legally acquired private property) in the first place, so I'm wholly unbothered by Google's actions and don't believe they reflect a moral wrong.
Clearing the forest and failing to get the permit are two distinct events here. Google is in trouble for failing to get the permit. Not for clearing the trees. Everyone on both sides was aware the trees were coming down. Think of it like camping at a campsite without getting a permit first.
Failure to get a permit is non-news for almost every entity on Earth. Except when you are megacorp, hence my comment.
Corporations should not break the law, period. Weird that anyone would defend that.
The forest was put up specifically to be cut down. The default state for these trees was to be cut down.
If you can’t see the difference between this and clear cutting of something like an old growth forest, then there’s nothing left worth saying to you.
There is limited public attention. Wasting it on such a minor infraction is unwise. Much better to focus on more egregious examples of corporate wrongdoing.
I’ve heard from multiple labs that model release is up to the US government now.
You can also see this with Astra being announced and then launched a few days later.
I imagine they all want to avoid uninformed knee-jerk legislation, so preventing a "three mile island" type incident that would freak out society and congress would be common ground alignment between all the labs.
The air around AI is already dense, but it would suffocate the industry overnight if something like a large scale financial attack or grid attack took place.
I don't buy it.
SpaceX is not exactly a small corporation, and they nevertheless strike me as far less risk-averse than GOOG and they deal with actual hardware that actually effing explodes.
If Google Cloud was in charge, Gemini 5 would be public, and they wouldn't be messing with Gemini 4 nor 1M context when they have 10M internal. C'est la vie.
Anyone can promise the moon in the future.
I use DeepSeek too but it’s more of a workhorse model, requires a lot more guiding and sometimes struggles to get where I want it.
If you compare to "smart" from a year ago, you have endless options that are basically free at this point which outperform everything from a year ago.
Deepseek Flash V4.1, Sonnet 5.5, GPT6 Luna, etc are really strong for the cost.
I thought it's not actually available to anyone? They announce huge challengers, but they don't exactly "push" them out...
“Helium” in Spanish is “Helio.” And Helios was the most powerful AI in the Deus Ex universe… Coincidence??
Does their Pro plan have similar limits to OAIs $100 business plan for example?
Any equivalent for the 20x plans?
Excited to try it, opus level benchmarks with their omniscience score and multimodality seems worth testing!
https://artificialanalysis.ai/articles/artificial-analysis-i...
https://news.ycombinator.com/item?id=50268868
https://theainewsreport.com/2026-10-28-artificial-analysis-i...
IMO it emphasizes the wrong things if we are interested in "intelligence"
There are many more aspects which affect outcomes
How do you do fellow frontier models? </meme>
Clicks link
Carbon isn't anywhere on the page and you can't toggle it on either. I'm sure it's just timing and it will be there soon but still just look amateur, which is about how I feel about Google and AI overall. I know the models have improved but I still have a bad taste in my mouth from testing Gemini in the past and I still feel that Google has an overly complicated "how to use our models" that changes way too often. That coupled with "you could lose access to your google account if we don't like what you do with Gemini" means I have little desire to test it out.
On top of all of that I'm completely over announcing a new model that we can't actually play with or try out. I don't think Google is lying about how well their model scored exactly but it still has big "my girlfriend goes to another school, you don't know her"-energy.
If you are going to hyperlink, maybe make sure that your model is actually on the page.
Also, the screen reader has a bug ([[duration]] minutes instead of the actual value) and you have once again announced a model we cannot use.
Edit: Both have been fixed after posting this. Maybe a Gemini 4 Carbon Agent has read and fixed the screen reader bug xD
I cannot fathom what motivates a lab, whose releases for the last year haven't been competitive, to shoot themselves in the foot like this. All this does is make it easier for competitors to snipe your actual launch, even if there was potential with this model (doubtful given 4 Argon, 3.1 Pro and 3 Pro), now everyone knows what exactly they need to release eval results wise to beat Carbon. There is this old idea of "Your competitor always is the first to have an idea what your next product will do", but in this case, Anthropic and co don't even need to guess, they got specific numbers to target.
Announce your models the moment they are available on API and in the subscription for all paying subscription users (cough Argon cough), stop with the cryptic tweets to drum up hype you know full well you won't live up to and make models that tool call, task adhere, task infer and handle long context tasks well beyond simple back-and-forth, chat style prompting. Reminder, during Ox Alpha, actual Google employees implied that was their model. How is this something anyone would believe is a good idea and how are they still allowed to tweet anything after? Please, stop that, it isn't helpful.
If frontier isn't your thing, stop wasting money on moonshot pre-train runs, fund research and open weight via Gemma and sell your inference hardware to other labs, that's far more lucrative anyway, as seen with Colossus 1.
[0] https://artificialanalysis.ai/evaluations/artificial-analysi...
So if they are still vetting and changes can thus happen, that'd mean the score could change for better or worse. Not sure they'd want the embarrassment of a #1 score changing after announcement.
Sort of? Yes, OpenAI has been bad about slow rolling but Anthropic's last few models have been available essentially at the same time as they announced it. In most cases I updated Claude Code and was using the new model right away.
Compare that to OpenAI who sometimes says "Available to all Plus/Pro/etc plans" without saying anything about "rolling out" and then it's a crap-shoot on when access will actually roll out. That said, they have gotten better, it used to be 1-3 weeks sometimes to get a new model/feature they announced was available for your plan.
And then there is Google... Who loves to talk about models like this and then give no timeline for release. Even when they do release the models it's miserable to find out how to even use it. Is it in the API? Which API? Google Cloud/Vertex/Gemini Cloud? Whatever it's called this week? Coding plan?
No, what Google does (or rather fails to do) in its rollouts does not compare to OpenAI/Anthropic.
As for the AA page, it's just embarrassing. Any smaller company doing that would get roasted, Google doesn't get a pass, learn how to coordinate or wait for something to actually be true/live before posting.
Gemini 4 Argon (at best Sonnet/Terra level) meanwhile took weeks to “come out” after their public announcement and when it finally did, paying Google AI plan users couldn't access Argon on Gemini or Antigravity even a week after. Three weeks from announcement to accessible for paying customers is not standard outside Google and this has been a problem with every Gemini release for the last year, so even using governmental requirements as an excuse isn’t possible here.
I will say that I dislike how OpenAI releases model only for Work and Codex but not via Chat, Anthropics "everywhere all at once" strategy is I feel the most user friendly.
It's on my screen. Try asking your agent to refresh the page.
Even in the optimistic case that this isn't some conspiracy, it highlights how arbitrary their scores are. I had trusted their previous ratings and their new index shows that even AA doesn't agree with them. What's to say this new index is any better and won't just change drastically again?
For historians, The blog post is 31 minutes ago so we can say that it was true for 20-30 minutes.
Slightly embarrasing but it seems to be the case that the model is actually worth what they were saying on AA. It's just that it is not available to people and I think that Anthropic and OAI will release a better model to public before google might.
AA has tons of its flaws too and the recent shift to their new v5 has been nothing but controversial for many.
Next: Silver, then Iodine
Slowly spelling out ARC-AGI
It has better tuning towards not adding unhelpful caveats and irrelevant details. When I told it I was unhappy with some slop being dumped into some repo markdown files from Astra and Sol it actually went ahead and tastefully cleaned them up in line with how I expected them to be. Reductions are normally so bad from agents.
One big thing is that it really feels like talking to someone smarter than me. Other agents just seem extremely knowledgeable but Argon felt smart.
If they were able to make flash 3.8 so good off of their 3.1 pro model, flash 4 will probably be really good. But ofc you'd want your pro model to be as strong as possible in coding so that flash can take all those gains and efficient-ize them.
It's why Opus 5.5 is a beast
Barium > Argon > Carbon
Why is carbon then more capable?
Any bad actor worth their salt is already a) going to have the weights downloaded, forever, to various models that are good at hacking and b) is out of reach of the U.S. government anyways. A ban on open weight models, plus paternalistic gatekeepers of cyber models like the big labs, just opens up Americans and American companies and municipalities to hacking from abroad. Or I guess to a really nice racket where only the big labs can offer a “we’ll continuously monitor and improve your cybersecurity for millions of dollars a year”.
“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.”
Obviously we want to be vetting the parties that get to use these weapons, because we want them to only be used as tools.
Mind you that Kimi in it's latest release nerfed cyber capabilities, as they stated. It doesn't make sense for China to be loose with this either.
I don't think I'm selfish, so that leaves me with naive or ignorant, I guess.
OpenAI: astral objects, bigger = better.
Anthropic: more pretentious = better.
Google: ¯\_(ツ)_/¯
But damn, is (talking about) releasing a new SOTA model two weeks after the previous one stupid. Do they not plan these things? Argon shouldn't be released at all if you're going to put out "argon 1.1" few weeks later.
When I mention a preference for brevity in a comment offhand and it snowballs into a rule 8 sessions down the road in a PR review that requests the removal of 90% of a colleague's docstrings.
I can't imagine all the "push back" on the "load bearing" Code of Conduct from their corporate foundation documents.
> [Header image is a chunk of ice]
???
What's next Krypton?
you can definitely tell new people are in charge at deepmind based on the change in naming
What happens when all the good names are taken from the periodic table of elements?
I have registered calcium, helium, and californium for my California-based LLM.
Silicon Valley Clowns...
Lol. The performance increase in the last 6 months has been massive.
Maybe I am just paranoid but at least if OAI or Anthropic ban my account I am not at the risk of losing access to my email address too.
You never know with Google and that is a problem.
I'm going to experiment with this one (assuming they ever get around to actually releasing it), and I'm really hoping the lame cultural aftertaste is gone.
Too bad you won't be able to use it to secure your open-source projects. If you try it'll throw up a prompt telling you to use some initiative you can't use or afford while the provider collects the findings. What a racket.
This post explicitly says that's been improved in Carbon. So, here's hoping. Google currently has the most generous usage limits on their $20 plan among US providers, so it's a pretty good deal.
Are you saying when you use Argon in antigravity it has generous token usage? I forget if I tried it with antigravity but my experience with using it as a chat bot was the usage limit was terrible.
The order of generosity, as it seems to be from my limited testing is currently Gemini->Claude->ChatGPT. (And, if we're including Chinese subscriptions, Kimi is below all of those. Very stingy.)
Come on googs!
With Anthropic and OpenAI, it's obvious immediately which are the better models. I.e. Sol is better than Luna, or Mythos is better than Fable.
But Google, let's pick Argon, which to most folks means mostly nothing. And then... Carbon! So which is better, Argon or Carbon? Nobody fucking knows. Plus neither sounds cool or viral or mysterious or even provokes curiosity.
When Fable was first leaked it sounded different and intriguing... Carbon just sounds... boring.
Oh yeah incredibly clear that a mythos is better than a fable /s
That's fairly obvious to me. Most people associate Myth with the gods, which are kind of the biggest and most powerful things we can imagine.
Not so much of a stretch...
Mythos > Fable ? Opus > Sonnet > Haiku
Tho I think it’s definitely not obvious at first glance. And requires some thinking to order if you are even familiar with the terms in the first place (I had to look up opus)
It is not obvious to me at all. I like Apple's naming more: iPhone Pro Max is better than iPhone Pro which is better than iPhone.
Sol vs Luna, how am I supposed to know which is better? Is sol a word, what does it mean? Same with mythos and fable. How am I supposed to keep track of these names across multiple companies? Why can't they use Claude vs. Claude Pro vs. Claude Pro Max?
The ordering seems to be based on size?
I actually like the google one. Alphabetical categories. Essentially alpha beta without the existing connation.
I think the bigger problem it’s getting harder to order models linearly. As later versions of the model make different tradeoffs across dimensions.
Mythos -> Fable -(sketchy)-> Opus -> Sonnet -> Haiku mostly checks out
Astra -> Sol -> Terra -> Luna checks out
Carbon ?? Argon
So no, OAI scheme is also confusing as hell. And they are going to add even more confusion when they release presumably Nova.
4 > 3.8 > 3.1 > 3.0
Unless you put Flash onto it, which denotes a lower cost version.
It seems like Carbon is a drop in replacement for Argon (same price). They can keep using random cutesey names as long as the version number + Flash naming scheme keeps making sense. Feels Android-y to me.
Wow. This huge. An AI company actually just said 'fabrication' instead of 'hallucination'. AI bubble about to burst...