Google opened Gemini 4 Carbon to paying developers and businesses on Tuesday, two weeks after first handing its strongest AI model to a small group of cybersecurity defenders.

The release puts Google’s most capable system within reach of anyone willing to pay for it, and it lands one day after rival Anthropic put its own flagship, Claude Fable 5.2, on sale. The two launches have turned November into a head-to-head contest over who builds the best model for long, complicated work — and who charges least for it.

Key Points

  • Gemini 4 Carbon went on general release on 10 November, after a two-week limited rollout to cybersecurity defenders through Google’s Fairwind programme.
  • Paying developers can call it through the Gemini API, and Ultra subscribers get it in the Gemini app, AI Mode in Search and Antigravity. Free-tier users are left out.
  • It costs $2 per million input tokens and $10 per million output tokens until 30 November, then doubles to $4 and $20 from 1 December.
  • Google is also previewing /swarm in Antigravity, which turns one prompt into a team of AI agents working in parallel.
  • Google says Carbon tops three workflow benchmarks, but public score tables reviewed by VentureBeat do not yet show its results for two of them.

What happened

Google announced Carbon on 28 October, calling it its strongest model, but kept it behind a gate: only selected cybersecurity defenders, vetted through its Fairwind programme, could use it. That gate came down on 10 November. The general-use version is now on the paid tier of the Gemini API in Google AI Studio and Android Studio, and in Gemini for business, Google said in a blog post.

Consumers reach it a different way. Google AI Ultra subscribers — the company’s top tier, priced at $99.99 a month for the standard plan and $199.99 a month for a higher-usage plan in the United States — are getting Carbon in the Gemini app, in AI Mode in Search and in Antigravity, Google’s coding platform. Google has not published a country-by-country rollout schedule, so some Ultra subscribers may still be waiting.

There is no free route in. The model is absent from the free tier of the Gemini API, and subscribers on the cheaper AI Pro plan keep Gemini 4 Argon, the frontier model Google released in October. Google describes Argon as the right choice for most software engineering and knowledge work, and Carbon as the model for jobs that push a system’s limits: multi-step coding and debugging, open-ended scientific research, and work that needs repeated checking.

Alongside the model, Google is previewing a feature called /swarm in Antigravity. One instruction becomes a coordinated team of agents that split a project apart and check each other’s work in real time. Developers can put Carbon on the hardest thinking — system design, coordination and auditing — while faster, lighter models handle the building in separate workspaces, with adversarial checkers testing every change.

Google also widened a programme that gives 10,000 scientists free access to its newest models in Antigravity for six months. Thousands have already joined, the company said, and Carbon is now included for current and future participants.

Why it matters

The price is the sharpest part of the launch. At $2 per million input tokens and $10 per million output tokens, Carbon undercuts its nearest rivals on paper: Anthropic charges $8 and $40 for Claude Fable 5.2, and OpenAI charges $10 and $50 for GPT-6 Astra. But the sticker price is not the whole bill. Carbon uses more tokens than Argon on hard jobs, thinking longer and running extra verification loops, and its introductory rate expires on 30 November. From 1 December the same tokens cost $4 and $20 — still half the price of Fable 5.2, but twice what early testers are paying now.

The benchmark picture is mixed, and it depends on which test you trust. Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR research test and Zapier’s AutomationBench, at more than twice the lower cost per task than the next best model. VentureBeat, checking the public tables behind two of those claims, found no Carbon result listed: Fable 5.2 leads the published CursorBench table, and WANDR’s table leaves Carbon blank. That does not prove Google’s numbers wrong, but it means they cannot be checked from the public record yet.

On the one broad ranking where every model is scored the same way, Carbon trails. The Artificial Analysis Intelligence Index puts Fable 5.2 at 65, Carbon at 60, Claude Opus 5.5 at 59 and GPT-6 Astra at 57. Drill into single jobs and the lead changes hands: Carbon scores 65.2% on Terminal-Bench 4.0, a test of terminal work, ahead of Fable’s 61.6%, while Fable wins FrontierCode 2.0, which grades changes to real codebases, 49.1% to 40.6%. The honest reading is that no single model leads everywhere, and the right pick depends on the job.

Carbon’s strongest card is security. A separate version of the model, built for cyber defence, stays inside the Fairwind programme and is not part of this public release. Google says that version leads tests for finding, exploiting and patching software flaws, and points to early fieldwork: the security firm Wiz used it to find previously unknown vulnerabilities at two large software providers in under two hours, work Google says would normally take weeks. The two companies were not named.

The model has also done striking work in research. DeepMind’s team used Carbon-powered agents to formally verify key security properties of the code behind Signal’s messaging protocol, generating millions of lines of mathematical proof in the Lean language — a job the company says has historically taken expert teams years. In drug research, the company Axiom Bio says Carbon is now its leading model for translational science, having found toxicity mechanisms its scientists had not spotted before. And on protein structure work, Carbon retrieved 40 times more evolutionary relatives than before across seven hard test families, giving structure-prediction tools far more to work with.

Google also published safety figures for the new model, all from its own testing. In simulated cyberattack trials, Carbon blocked 99% of attack attempts, and the breach rate per attempt fell to 0.8%, down from 3.1% for Argon. External testing by Trajectory Labs found no new universal jailbreak across 132 hours and more than 13,000 requests, and on a 375-task honesty test the model passed 92% of tasks, against 70% for Argon. It also scored zero successful prompt-injection attacks on one industry benchmark — a first, according to Google, though the company notes that prompt injection remains an area it is still working to improve.

What happens next

The test that matters now is not a benchmark but a bill. Teams that adopt Carbon in the next three weeks get the introductory price; from 1 December they pay double, and will learn whether the model’s extra thinking earns its keep or simply burns more tokens per finished job. Google’s own advice points the same way: keep the current model as a baseline, give both systems the same work, and compare what each one costs to finish.

The robot version of the same technology is already on factory floors, with Gemini Robotics 2.5 running a humanoid on live Mercedes-Benz lines. Carbon’s computing power, meanwhile, rests on Google’s newest AI chips, unveiled earlier this month. And the rivalry is not standing still: Anthropic’s Claude Fable 5.2 arrived the day before Carbon, and already leads the broad index Carbon once topped.

For now, Carbon’s opening is a statement of intent as much as a product launch: Google is betting that the model which does the hardest, longest jobs — and does them cheapest per task — will win the next round of the AI race, even if it does not win every test along the way.

Sources: Google blog, 10 November 2026; VentureBeat, 10 November 2026; The Verge, 10 November 2026; BusinessToday, 11 November 2026. Published 11 November 2026.