Gemini 4 Carbon, Google's strongest model, is now open to the public. Google announced it on 10 November 2026, two weeks after Carbon first went to a small group of cyber defenders. That's the good news. The bad news, for me at least: on the consumer side it's only in Google AI Ultra, Google's most expensive plan. I'm on Google AI Plus, so I can't use it in the Gemini app, and I couldn't test it for this post. Below is what Google says Carbon can do, who can use it and what it costs. All figures come from Google's own announcement, model card and pricing page.
What's New Since Carbon Was Announced?
When I first wrote about Gemini 4 Carbon's benchmarks on 29 October, nobody outside the cyber-defence pilot could use it and there was no date. Now there is:
- Developers can use it through the paid tier of the Gemini API, in Google AI Studio and Android Studio. There's no free tier: Google's pricing page says "Gemini 4 Carbon is not available on the Free tier."
- Businesses get it in Gemini Enterprise, Google Antigravity and CodeMender.
- Consumers on Google AI Ultra get it in the Gemini app, AI Mode in Search, Antigravity and AI Studio.
- Scientists can apply for six months of free access to Argon and Carbon in Antigravity's highest usage tier.
- Cyber defenders can still apply to the Fairwind Program, which gets a version with fuller cybersecurity skills.
The model card adds a few basics that weren't public in October: a context window of up to 1 million tokens, input as text, images, audio and video, and a knowledge cutoff of March 2026. Google describes Carbon as built on Gemini 4 Argon, using more "thinking" tokens to explore, reason and check its own work.
Can I Use Gemini 4 Carbon on Google AI Plus or Pro?
No, not in the Gemini app. Google's AI plans page (checked 11 November 2026, Australia) splits the models like this:
| Plan | Top Gemini model listed | Gemini 4 Carbon? |
|---|---|---|
| Google AI Plus (my plan) | Gemini 3.6 Flash and Deep Research | No |
| Google AI Pro | Gemini 4 Argon | No |
| Google AI Ultra (5x and 20x) | Gemini 4 Carbon and Argon | Yes, in the Gemini app, AI Mode, Antigravity and AI Studio |
The page's own wording is "Available to Google AI Ultra subscribers in selected countries." Google hasn't said whether Pro or Plus will get Carbon later. So for now, if you want Carbon without Ultra, the realistic route is the paid API, where you pay per token.
How Much Does Gemini 4 Carbon Cost?
Through the API it starts cheap, then doubles. Until 30 November 2026 Carbon costs US$2 per million input tokens and $10 per million output tokens, the same as Argon. From 1 December the price goes to $4 and $20, which is exactly what Claude Opus 5.5 costs, and half of Claude Fable 5.2's $8 and $40. Batch jobs are half price, and cached input is $0.10 per million (then $0.20).
One thing to watch: output prices include thinking tokens, and Google says Carbon "may use more tokens to maximize performance, especially at higher thinking levels." So the real cost of a task can be higher than the per-token price suggests. I saw the same problem with Claude Fable 5.2, which drained my Pro limit fast.
What Does Google Say Carbon Is Good At?
Google pitches Carbon for the hardest work: long, multi-step coding and debugging, research, and cybersecurity. Google says Argon is still the better pick for most everyday coding and knowledge work. The new claims in this announcement:
- Coding agents. First place on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2 times lower cost per task than the next best outside model.
- Knowledge work. Second on the new Multimodal Knowledge Agent (MKA) benchmark from Google and Artificial Analysis, but more than 6 times cheaper per task than the leader.
- Teams of agents. A preview of
/swarmin Google Antigravity, which turns one prompt into a team of agents. Carbon plans and checks the work while faster models do the parallel parts. - Security. Google says Carbon helped show that two bugs in popular open-source software were exploitable, and that Wiz used it to find zero-day bugs at two big SaaS companies in under two hours.
- Science. In a protein test across 7 protein families, Carbon found 40 times more related proteins on average without retraining the structure models. Google says wider testing is still going on.
The full benchmark table hasn't changed since October (Google labels the results "as of October 2026"), so for the head-to-head numbers against GPT-6 Astra and Claude Opus 5.5, see my first Carbon post.
Is Gemini 4 Carbon Safer Than Argon?
On Google's tests, yes. These numbers are new in this announcement:
- In simulated cyberattacks, Carbon blocked 99% of attack paths from an automated red-team agent. The breach rate per attempt fell to 0.8%, down from 3.1% for Argon.
- In Gray Swan's testing, automated cyberattack requests succeeded 2.17% of the time, down from 9.86% for Argon.
- Trajectory Labs found no new universal jailbreaks in 132 hours and more than 13,000 requests.
- On a 375-task honesty test, Carbon passed 92% of tasks against Argon's 70%. It refused to make up missing experiment data 98% of the time.
Google is also honest that it's not finished: it says resistance to prompt injection "still needs work". As always, these are Google's own tests.
What Do I Make of It?
It's too bad. In October, the problem was that nobody could use Carbon. Now the problem, for me, is my plan. I'm on Google AI Plus, and Carbon is only in Ultra on the consumer side, so I couldn't give it the hands-on test I promised. Even Gemini 4 Argon, the model below it, is listed for Pro, not Plus.
I understand the logic. Carbon uses a lot of thinking tokens, and Google is still careful about its cyber skills. But it does mean the comparison I most wanted, Carbon against Claude Opus 5.5 and Fable 5.2 on my own website, will have to wait. The cheapest way to try it is the API before 30 November, when the intro price ends. Until then, the numbers above are Google's, not mine.
If you're on Ultra and have tried Carbon, I'd like to hear how it compares with Argon. You can reach me on the contact page.
Sources
- Google: Gemini 4 Carbon is now available by Koray Kavukcuoglu, 10 Nov 2026: availability, pricing, new benchmark claims, demos and safety results.
- Google DeepMind: Gemini 4 Carbon model card (context window, inputs, knowledge cutoff, distribution) and evaluation methodology.
- Google: Gemini API pricing and Google AI plans, both checked 11 Nov 2026.
- Claude Opus 5.5 and Fable 5.2 prices: from Anthropic, see my Fable 5.2 post.
Checked 11 November 2026. All benchmark and safety figures are Google's own, and I haven't been able to verify them. The images were made locally with Qwen Image 2.1, the chart was redrawn from the official prices, and AI helped me research and edit this post.
