Google has made Gemini 4 Carbon, its most capable AI model, generally available to developers and paying subscribers — two weeks after first announcing it, and one day after Anthropic launched its rival Claude Fable 5.2. The headline for buyers is the price: Carbon starts at $2 per million input tokens and $10 per million output tokens, a quarter of Fable 5.2’s $8 and $40 — though only until November 30, when Google doubles the rate.
Carbon was unveiled on October 28 as Google’s strongest model, built for long, multi-step engineering work and for cybersecurity defence. Until now, almost nobody could use it. Access was limited to a small group of cyber defenders through Google DeepMind’s Fairwind programme, with everyone else told to wait for a date Google had not set. That wait ended on Monday, when Google announced that developers can now build with Carbon through the paid tier of the Gemini API in Google AI Studio and Android Studio.
Who can use Carbon now
For developers, Carbon is available through the paid Gemini API in AI Studio and Android Studio, and through Gemini for business. For consumers, it is rolling out to Google AI Ultra subscribers — the $99.99-a-month Ultra 5x tier and the $199.99-a-month Ultra 20x tier in the US — in the Gemini app, in AI Mode in Search, and in Antigravity, Google’s agent-first development platform.
Google has not published a country-by-country rollout schedule, and it has not said when — or whether — Carbon will reach Vertex AI, the cloud platform many enterprises buy through. The model is proprietary and hosted only: there are no downloadable weights and no self-hosting option.
The price, and the catch
The introductory price of $2 per million input tokens and $10 per million output tokens runs through November 30. After that, Carbon costs $4 and $20 — still half of Fable 5.2’s standard rates, and the same as OpenAI’s GPT-6 Astra at its post-introductory level. Google also offers a 95 percent discount on cached input tokens, which matters for teams that repeatedly send the same code repositories or documents: during the promotion, cached input works out at about $0.10 per million tokens.
Price per token is only part of the bill, though. What matters is what a finished job costs, including retries and corrections. On that measure the picture is mixed. On Zapier’s AutomationBench, which tests whether agents can complete assignments across business apps, Carbon scores 53.88 percent at an estimated $2.12 per task, ahead of Fable 5.2 at 51.45 percent and $4.64. But on Artificial Analysis’s broad index, Fable 5.2 now leads at 65 points to Carbon’s 60, after Carbon debuted in first place at 59 when it was announced.
What Carbon is built for
Google pitches Carbon at sustained agentic work: engineering problems that take days rather than minutes, where a model has to persist through dead ends, find the root cause of a subtle bug across a large codebase, and patch it. Internally, Google says its teams have seen Carbon run for days on such tasks. It also claims a jump of 89 elo points in web development, to fourth place on the WebDev Arena leaderboard.
The sharper claim is in cybersecurity. Carbon scores 88.2 percent on SEC-bench Pro for finding real vulnerabilities in software, 54.8 percent on ExploitGym for demonstrating they can be exploited, and 71.0 percent on CWE-bench for patching them — all state-of-the-art results by Google’s account, though the tests were run without guardrails and the headline numbers are Google’s own. Google says its Chrome team found Carbon more than ten times better than another lab’s large frontier model at generating high-quality patches, and that the security platform Wiz used Carbon to find previously unknown vulnerabilities while scanning two major software providers in under two hours.
Alongside the launch, Google is previewing a feature in Antigravity called /swarm, which lets a team of AI agents divide up a project and check one another’s work, with Carbon handling architecture, coordination or audit tasks while lighter models take parallel implementation jobs.
A crowded week at the frontier
Carbon’s general availability lands in the middle of a crowded fortnight. Anthropic launched Fable 5.2 and Mythos 5.2 on November 9, Meta launched its smallest model, Muse Ember, the same day, and NVIDIA unveiled its GR00T 2 robot model. Google’s answer is not to outscore the field on every test — Fable 5.2 leads Carbon on several coding and science benchmarks, including FrontierCode 2.0 — but to undercut it on price while leading on the tests closest to real agentic work.
For teams deciding whether to try it, the honest test is a pilot: give Carbon and the current model the same jobs, count the tokens, the retries and the reviewer time, and work out what the same run costs at December’s rates. The introductory price expires on November 30, which gives buyers about three weeks to find out whether Google’s cheapest frontier model earns a place in their stack.
