The Black Cloud

AI - Google Gemini 4 Carbon is #1, but can you actually use it? 🤣

Google has barely finished talking about Gemini 4 Argon, and now we have Carbon 🤣 This time, it has taken the top spot on the updated Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 and GPT-6 Astra.

There is just one small problem: most of us still cannot use it.

Gemini 4 Carbon

Carbon is Google’s latest frontier model, focused on autonomous coding, cybersecurity defence and sustained, multi-step work. The interesting part is the whole workflow: finding vulnerabilities, checking whether they can be exploited and helping to fix the underlying code. Google also reports internal engineering tasks running for days.

Official announcement

Gemini 4 Carbon: our next frontier model

NOTE
As of 29 October, Gemini 4 Carbon is not generally available. Google is starting with a subset of cyber defenders, then extending access to trusted partners through its Fairwind Program. It is also participating in the US government’s voluntary pre-release testing process. There is no confirmed date for wider access or list of consumer plans that will include Carbon.

What is new?

The main changes in Google’s announcement are:

My take: Google is back, but not for me 😭

Let me be fair first. If Carbon turns these results into useful everyday work, this is a big deal. More competition for Anthropic and OpenAI is good for the rest of us.

I understand the caution around a model with stronger cybersecurity capabilities. But being shown Google’s best model while waiting to find out whether I can use it still feels a bit like looking at a new car through the showroom window 🤣

It is even more annoying after Google’s changes to Pro model access. AI Plus subscribers are losing selectable Pro access in Gemini Apps, while AI Pro and AI Ultra retain it. I pay for AI Plus, so another impressive announcement without knowing what my subscription will actually get is rather frustrating 😭

One important distinction: Google’s current support page says free accounts using Auto can still have some harder prompts routed to Pro. That is not confirmation of Carbon access for free users or AI Plus subscribers.

Not available yet

The reactions sum it up quite well: plenty of excitement, followed by people asking when they can actually try it.

The frontier is fast, and the internal vibes on this model are good. Rolling out to cyber defenders first, then more widely available as the next step!

Josh Woodward confirms the order of the rollout, but there is still no date for the rest of us.

I would love to try it out, Paige. That is rather the problem 😉

X Freeze is making predictions about adoption and compute, not announcing which subscriptions Google will include. I would not take access for granted.

Benchmarks and rankings

Google has a result worth celebrating, but I would keep the different leaderboards separate.

First on the Intelligence Index does not mean first at every job. Artificial Analysis places Carbon behind Claude Opus 5.5 on its Coding Agent Index. Arena’s launch results put Carbon first for text, second for agents and fourth for web development. Test setups and reasoning settings matter too, so the numbers below are not all interchangeable comparisons.

Is Gemini underrated?

IF THIS WERE A NEW ANTHROPIC MODEL, THE ENTIRE TIMELINE WOULD BE LOSING ITS MIND 💀

That is a very interesting take, and probably true 🤣 Google deserves credit for the result. I still want to see how Carbon behaves in real work, but dismissing it simply because it is Gemini would be rather silly.

Cybersecurity and alignment

This is the part that makes the slower rollout understandable. Finding vulnerabilities is useful; checking the risk and helping to fix the underlying code makes the capability much more consequential.

The distinction between internal testing and external benchmarks matters here. So does Google’s disclosure that its cyber benchmark results were evaluated without guardrails.

Google’s 0.0% prompt-injection result means none of the attacks in that Gray Swan test succeeded. It does not mean Carbon is impossible to jailbreak or safe in every environment.

There is a useful counterpoint on hallucinations too. Artificial Analysis measured 11% for Carbon versus 4% for Argon on AA-Omniscience v1.1. Carbon still has the second-lowest rate in that comparison, but these are results from one evaluation, not the odds that any answer you receive will be wrong.

Long-horizon coding

Running on a task for days is the claim I would like to see tested most carefully. Kenneth, who works at DeepMind, describes his internal experience:

Carbon is writing all of my code and I can trust it to run all night, for multiple days, making intelligent decisions autonomously.

That sounds excellent. I would still want independent testing before treating an overnight run as a substitute for reviewing the work.

Chubby’s recursive self-improvement interpretation goes further than the announcement. Google says its researchers and models are working together to accelerate progress; that is not the same as demonstrating autonomous recursive self-improvement.

3D and spatial reasoning

The 3D demos are fun to look at, although a few impressive examples are not enough to settle which model is best.

Haider highlights Blueprint-Bench 2.0, which tests reconstructing floor plans from interior photographs. That is a specific spatial task, not proof of human-level understanding across the board.

Pricing

Small but important detail: the $2 input/$10 output per million tokens price is introductory. Google’s footnote gives a standard price of $4/$20 after that period, with cached input tokens discounted by 95% against the input rate.

Artificial Analysis uses the standard pricing for its comparisons. Even with lower per-token costs, a model running for days can use a lot of tokens. I care more about the cost of the finished job than an attractive headline token price 😉

And finally…

Impressive

Elon Musk congratulating Google. Did not see that one coming 🤣

<< Previous Post

|

Next Post >>

#Blog #AI #Google #Google Deepmind #Gemini 4 #Gemini 4 Carbon #Gemini 4 Argon #Cybersecurity #Benchmarks