Gemini 4 Carbon, Google’s Most Capable Model, Is Now Available to Developers and Ultra Subscribers
Two weeks after restricting it to cyberdefense specialists, Google is opening Gemini 4 Carbon to paying developers and enterprises via the API, as well as to Google AI Ultra subscribers in the Gemini app, Search's AI mode, and Antigravity. Launch pricing is set at $2 and $10 per million tokens through November 30, before doubling.
Gemini 4 Carbon is no longer reserved for cyberdefense specialists. On Tuesday, November 10, around 5:00 PM Paris time, Google opened its most capable model to paying developers and enterprises via the API, as well as to Google AI Ultra subscribers in the Gemini app, Search’s AI mode, and the Antigravity development environment. When it was announced on October 28, Carbon was only accessible to a small group of cybersecurity defenders, with no release date specified.
Who Gets Access, and at What Price
According to the blog post by Koray Kavukcuoglu, Google’s Chief AI Architect, Carbon is available:
- to developers on the paid tier of the Gemini API, in Google AI Studio and Android Studio;
- to enterprises in Gemini for business (via the API or the model menu), Antigravity, and CodeMender;
- to the general public subscribed to Google AI Ultra, in the Gemini app and Search’s AI mode;
- to selected scientists, free for six months in Antigravity, as an extension of a program launched four weeks ago for Argon.
Access to “Carbon for cyber,” the variant with the most advanced offensive capabilities, remains tied to the Fairwind program by application.
On the API side, the pricing table confirms the rates announced in late October: $2 per million input tokens and $10 per million output tokens, including thinking tokens, through November 30, then $4 and $20 starting December 1. Cached inputs cost $0.10 per million, rising to $0.20. The model is not available in the API’s free tier. Google does not state whether Carbon will ever arrive in subscription tiers below Ultra.
Argon for Everyday Tasks, Carbon for the Hardest Problems
Google is not positioning Carbon as a replacement for Argon. Argon remains, according to the company, “the ideal choice for the majority” of coding and productivity tasks. Carbon targets jobs that “test the limits” of a model: multi-step debugging, architectural refactoring, and open-ended scientific research. It consumes more tokens than Argon to explore, reason, and verify its work. At equal unit pricing during the promotional launch period, the same task will therefore often cost more with Carbon.
For complex projects, Google is previewing /swarm in Antigravity: a single prompt kicks off a team of coordinated agents, monitored in real time. Carbon can take on the most demanding roles (design, coordination, auditing) while lighter, faster models execute in parallel across isolated environments.
Strong Benchmark Showings, Mostly Measured by Google
Google claims that Carbon ranks 1st on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, with a cost per task “more than two times lower” than the best external model. Cursor confirmed this afternoon that the model is now available in its editor, where it scores 60.4% on CursorBench at high effort—the highest score on Cursor’s internal benchmark. On the Arena homepage, checked on November 10 around 6:00 PM, Carbon sits 1st in the Text category at High effort in the new model rankings; these leaderboards rely on user votes and shift frequently.
The blog post also highlights internal use cases: Google researchers report using Carbon, alongside a new tool called AlphaProof Code, to mathematically prove security properties of Signal’s protocol code, while the team at Wiz claims to have uncovered novel (zero-day) vulnerabilities across two major enterprise cloud software vendors in under two hours. None of these findings have yet been published in detail.
Security, According to Google’s Metrics
Because Carbon is more capable in cybersecurity, Google says it has strengthened its safeguards. In simulated attack environments, the model blocked 99% of trajectories driven by an adaptive attack agent, and the breach rate per attempt dropped to 0.8%, compared to 3.1% for Argon. On an internal 375-task agent honesty test, it succeeded on 92%, versus 70% for Argon. These figures come from Google and partners it selected; a detailed safety report accompanies the release.
Our take
Google stuck to a tight schedule: thirteen days between the restricted announcement and the paid rollout, whereas Anthropic still limits Mythos 5.2, its most comprehensive version, to vetted organizations. The key issue for engineering teams is pricing: at $2 and $10 per million tokens, Carbon debuts at the exact same rate as Argon, but this launch pricing only lasts twenty days, and Google itself cautions that the model generates more tokens than Argon. The true cost will have to be evaluated per task, not per token, and over time. Early external signals are promising (Cursor, the Arena), but they remain incomplete: CursorBench is proprietary to Cursor, and Arena votes measure user preference rather than reliability over long-horizon tasks. For the general public, the launch remains walled inside the most expensive subscription: for most Gemini users, nothing changes yet.
Sources
- Google — Gemini 4 Carbon is now available (blog post by Koray Kavukcuoglu, November 10, 2026) Primary source
- Google AI Studio — Gemini API pricing Primary source
- Sundar Pichai (X) — Gemini 4 Carbon rollout Primary source
- Google AI (X) — Gemini 4 Carbon available Primary source
- Cursor (X) — Gemini 4 Carbon available in Cursor Primary source
- Arena — Leaderboard (accessed November 10, 2026) Primary source
- The Verge — Google releases Gemini 4 Carbon with agent ‘swarm’ feature Secondary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
