Google’s strongest model just left the locked room.
Gemini 4 Carbon was announced two weeks ago, but only a few cyber defenders could use it. Now it’s open to paying developers and top-tier subscribers. Here’s what that means, what it costs, and where the hype gets ahead of the data.
Quick answer
What happened? On November 10, 2026, Google made Gemini 4 Carbon broadly available. It was first announced on October 28 with limited access.
Who can use it? Paying Gemini API developers, Gemini for business customers, and Google AI Ultra subscribers. Not free users.
What does it cost? $2 per million input tokens and $10 per million output tokens until November 30. Then $4 and $20.
Is it the best model? Not across the board. It leads on some tests and trails Claude Fable 5.2 on others.
What is Gemini 4 Carbon?
Google calls it its strongest model, with frontier cybersecurity skills and long-horizon agent work. That means tasks where a model plans, acts and self-corrects over many steps.
It sits above Gemini 4 Argon. It doesn’t replace it. Google says Argon is still the better pick for most software engineering and knowledge work. Carbon spends more tokens reasoning and checking, so it suits harder jobs. If you missed the earlier launch, read my Gemini 4 Argon explainer.
Artificial Analysis lists a 1 million token context window and says Carbon takes text, images, speech and video, and outputs text. Google’s own post doesn’t state those specs. It’s also proprietary, with no open weights.
Where can you use it?
- Gemini API: the paid tier, through Google AI Studio and Android Studio
- Gemini for business: API and model picker access
- Google AI Ultra: Gemini app, AI Mode in Search and Antigravity, rolling out. Ultra starts at about $100 a month in the US.
- Scientists: free six-month access to Argon and Carbon through Antigravity’s top usage tier. You apply, and eligibility and capacity limits apply. Engadget reports 10,000 scientists.
Engadget’s headline says “worldwide,” but the article names no countries. VentureBeat says the API is in “supported countries and territories.” Google hasn’t published a country-by-country schedule, and Vertex AI availability isn’t specified.
How much does it cost?
Per million tokens, standard uncached rates, according to VentureBeat:
- Gemini 4 Carbon, until Nov 30: $2 input, $10 output
- Gemini 4 Carbon, after: $4 input, $20 output
- Claude Opus 5.5: $4 input, $20 output
- Claude Fable 5.2: $8 input, $40 output
Cached input is 95% cheaper. That’s Google’s announced discount. Cheap tokens don’t always mean a cheap bill, though. Retries and human fixes add up.
How does Carbon compare with Claude Fable 5.2?
It depends on the test. Here’s the mixed picture.
Artificial Analysis Intelligence Index: Fable 5.2 scores 65. Carbon scores 60. Opus 5.5 scores 59 and GPT-6 Astra 57. So Carbon isn’t number one there.
Where Carbon leads:
- Terminal-Bench 4.0: 65.2%, against 61.6% for Fable’s best listed run
- Zapier’s AutomationBench: 53.88%, against 51.45% for Fable at max effort
- Cost per task on the Artificial Analysis suite: about $7.76, against $26.54 for Fable at max effort
Where Fable leads:
- FrontierCode 2.0: 49.1% against 40.6%
- FrontierSWE v2: 69.9% against 61.2% (reported by Anthropic)
- Terminal-Bench-Science: 74.6% against 64.8% (reported by Anthropic)
Google also claims first place on CursorBench 4.0, Perplexity’s WANDR and AutomationBench, at more than twice lower cost per task. VentureBeat notes the launch materials omit some scores, and public tables don’t show a Carbon row for CursorBench or WANDR. Treat those claims as unproven for now.
What about the cybersecurity claims?
Google’s October tests report 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting) and 71.0% on CWE-bench (patching). Google ran them without guardrails.
Here’s the catch. The cyber-specialized version, with fewer guardrails, is only available to selected defenders through the Fairwind Program. Google’s post doesn’t say whether the public Carbon matches the variant tested in those cyber evaluations.
Google also says Wiz used Carbon to find zero-day flaws in two major SaaS providers in under two hours. Google’s own numbers say successful attacks on Carbon fell versus Argon, and an outside group found no new universal jailbreaks in 132 hours. Google also says robustness against prompt injection still has room to improve.
What is /swarm?
It’s a preview feature in Google Antigravity that coordinates multiple agents from a single prompt. It’s a preview. Google hasn’t said when it becomes generally available.
Should you switch?
- Use Argon for everyday coding and knowledge work. Google itself recommends it.
- Try Carbon for hard multi-step engineering, security-adjacent analysis and long-running agent tasks.
- Move fast on price. The $2 and $10 rate ends November 30. Plan for $4 and $20 after that.
- Test on your own tasks. Benchmarks are mostly self-reported or mixed. Your workload is the real test.
- Compare cost per finished task, not just price per token.
What we still don’t know
- Parameter count, architecture and training data
- Independent verification of Google’s cost and safety figures
- Exact country and Vertex AI availability
- Whether the public model matches the cyber variant tested
- Usage limits for Ultra subscribers in the Gemini app
FAQ
When did Gemini 4 Carbon become available?
November 10, 2026. It was announced October 28 with limited access.
How much does Gemini 4 Carbon cost?
$2 input and $10 output per million tokens until November 30. After that, $4 and $20.
Who can use Gemini 4 Carbon?
Paying Gemini API developers, Gemini for business users and Google AI Ultra subscribers. Scientists can apply for free access.
Is Carbon better than Claude Fable 5.2?
Not overall. Fable 5.2 scores higher on the Artificial Analysis index. Carbon leads some tests and costs less per task.
Is Carbon better than Gemini 4 Argon?
For hard tasks, Google says yes. For most daily work, Google recommends Argon.
Is Carbon free?
No, except for scientists who apply and are accepted.
Can I download Carbon?
No. It’s proprietary and hosted by Google.
Bottom line
Carbon is a strong model at a cheap launch price. It isn’t the clear winner. It trades blows with Claude Fable 5.2 and costs far less per task on some tests.
Take the intro price, run your own tests, and don’t trust any single leaderboard. And mark November 30 on your calendar.