Google announced on October 28, 2026 Gemini 4 Carbon, which leads the Artificial Analysis Intelligence Index v5 with 59 points. The model is currently only open to cyber defenders, then to partners of the Fairwind Program, with no set date for developers.
Access by Tiers, No Date for Developers
The announcement post, signed by Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect at Google, describes this gradual access. Carbon is open from launch week to “a subset of cyber defenders,” then to “other trusted partners” via the Fairwind Program. Google states it continues “to gather feedback and adjust safeguards” before opening the model “as soon as possible” to developers, businesses, and the general public, with no specific date. On October 30, the list of models in the Gemini API only included one Gemini 4, Argon, released on October 13.
The post indicates that Google is participating in the “voluntary U.S. government process” for testing models before release, without naming the agency. The Center for AI Standards and Innovation (CAISI), part of NIST, announced on May 5, 2026 agreements with Google DeepMind, Microsoft, and xAI to evaluate models “before they are publicly accessible.” Google does not say if this is the mechanism, and NIST does not describe these agreements as voluntary.
Top Spot with a One-Point Lead
Artificial Analysis published the fifth version of its index on October 27. That day, Claude Opus 5.5 was leading with 58 points, ahead of GPT-6 Astra at 57. On the ranking viewed on October 30, Carbon, in its “high” reasoning setting, reaches 59 points, one more than Opus 5.5 and two more than GPT-6 Astra, the latter two measured in their maximum settings. Its entry ranks it first among 57 comparable models. A one-point lead on an index that aggregates nine evaluations remains slim, and Artificial Analysis does not publish a margin of error to indicate it is significant.
Google also claims records on security tests of its own choosing, including 88.2% on SEC-bench Pro, dedicated to vulnerability research. These measurements were taken “without safeguards,” while the company is still adjusting these protections before any broad release.
Carbon is not the only model reserved for a limited circle. Anthropic has reserved Claude Mythos Preview since April 7, 2026 for members of Project Glasswing, and the company writes: “we do not plan to make Claude Mythos Preview generally available.” Mistral AI, for its part, reserves a variant of Mistral Large 4 specialized in cybersecurity for partners like Orange Cyberdefense. None of these models have a general release date.
Introductory Pricing Set to Double
Google is launching Carbon at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. A token is a unit of text the model reads or produces. A note in the post warns that this pricing will increase to $4 and $20 upon the expiration of the introductory period, without dating it. For Argon, the date was specified: the price doubles on December 1, after an introductory period ending on November 30, 2026. The pricing page already displays both rate grids.
At full price, Carbon will cost the same as Claude Opus 5.5, billed at $4 and $20 according to ActuIA’s API price list, and cheaper than GPT-6 Astra, at $10 and $50. On October 30, the Fairwind Program page, which boasts “over 650 partners,” was still offering Gemini 4 Argon, in a version with “cyber defense capabilities.” Carbon was not listed.
