As of October 30, the Gemini API model list has only one Gemini 4, Argon, released on October 13. Gemini 4 Carbon, which Google presented two days earlier as its “most powerful” model, is not there, and the pricing page does not devote any line to it. However, it is first in the Artificial Analysis Intelligence Index v5, with 59 points. Google does not sell it to any self-service company. Artificial Analysis says it is “available via API from a provider”, Google’s API, which Google’s documentation does not describe.
Cyber defenders first, developers without date
The announcement post, signed October 28 by Koray Kavukcuoglu, senior vice president of Google DeepMind and Google’s chief AI architect, describes access in circles. Carbon is open from launch week to “a subset of cyber defenders”, then will gradually be opened to “other trusted partners” through the Fairwind Program. Developers, businesses and the general public will come next: Google says it will continue “to collect feedback and adjust safeguards” before opening the model to them “as soon as possible”. No date is given.
The same post indicates that Google participates in the “US government’s voluntary process” of testing models before publication, without naming the organization. The Center for AI Standards and Innovation (CAISI), attached to NIST, announced on May 5, 2026 agreements with Google DeepMind, Microsoft and xAI to evaluate models “before they are accessible to the public”. Google doesn’t say if this is the process it’s talking about, and NIST doesn’t call these agreements voluntary.
One point ahead of Claude Opus 5.5
Artificial Analysis released the fifth version of its index on October 27. That day, Claude Opus 5.5 led with 58 points, ahead of GPT-6 Astra at 57. It lost first place the next day. On the ranking consulted on October 30, Carbon, in its “high” reasoning setting, takes the lead with 59 points, one more than Opus 5.5 and two more than GPT-6 Astra, both measured in their maximum setting. Its profile ranks it first out of 57 comparable models. A point of difference on an index which aggregates nine evaluations remains slim, and Artificial Analysis does not publish a margin of error which would allow it to be considered significant.
Google also claims records on security tests chosen by it, including 88.2% on SEC-bench Pro, dedicated to the search for vulnerabilities, and specifies that these measurements were carried out “without safeguards”. These are precisely the protections that the company is still adjusting before any wide opening.
Anthropic and Mistral also keep closed circle models
On October 27, Anthropic’s Frontier Red Team described in a post critical flaws found by “an internal, unpublished research model” in browsers and operating systems. In the Chrome bulletins, an Anthropic researcher signs one flaw out of ten corrected in October, without being able to link these credits to this model. The post says nothing about a possible commercialization of this model.
Claude Mythos Preview follows a more explicit logic. Since April 7, 2026, Anthropic has reserved it for members of Project Glasswing: Amazon Web Services, Apple, Google, Microsoft and seven other launch partners, then more than forty organizations that maintain critical software. The page writes: “we do not plan to make Claude Mythos Preview generally available, but our goal is, ultimately, to allow our users to securely deploy Mythos class models on a large scale.” After the research phase, participants will pay $25 per million tokens in and $125 out.
Mistral AI, which published the weights of Mistral Large 4 on October 29, reserves a specialized cybersecurity variant for partners like Orange Cyberdefense. His post subjects this preview access to “rigorous verification and contractual safeguards,” with testing limited to owned or explicitly authorized systems. None of these four models have a general opening date. Carbon is the only one to occupy first place in a general index.
A call rate that will double, like that of Argon
Google is launching Carbon at the same introductory price as Argon: $2 per million tokens in, 10 out, and a 95% discount on the entry served from the cache, or $0.10. A note at the bottom of the ticket warns that when the introductory period expires, the price will increase to 4 and 20 dollars. The note does not date this expiration. Artificial Analysis already displays Carbon at 4 and 20 dollars, based on Google’s API.
Argon shows how the mechanism works. The note on the Argon ticket dated this expiration: November 30, 2026. The Gemini API price page already displays the two prices, one under the other: 2 and 10 dollars until November 30, 4 and 20 dollars from December 1. The price of Argon therefore doubles seven weeks after its launch. Argon was open to developers on October 13; its call price applies throughout this period.
At full price, Carbon will cost as much as Claude Opus 5.5, billed at $4 and $20 according to ActuIA’s API price list, and two and a half times cheaper than GPT-6 Astra, at $10 and $50. The only price on which a company can build a Carbon budget is that of 4 and 20 dollars: Google does not date the opening of the model to customers nor the end of its introduction period.
On October 30, the Fairwind Program page, which claims “more than 650 partners”, still offers Gemini 4 Argon in a version with “cyber defense capabilities”, as well as the CodeMender code correction agent. Carbon is not there.
Quotes translated from English by ActuIA.
Our articles will be prioritized in Google’s “Featured” section.
