Google sets Gemini 4 Carbon prices ahead of restricted access

Google has announced introductory prices for Gemini 4 Carbon, a superintelligence (SI) model it plans to offer first to selected cybersecurity defenders this week. Its October 28 announcement sets out a restricted rollout, not a public release, with wider access to follow through its Fairwind Program.

Carbon follows Gemini 4 Argon, which Google announced on September 30, 2026, while rolling it out to trusted cybersecurity defenders through Fairwind. Google’s October 28 Carbon announcement described Argon’s launch as two weeks earlier.

Carbon’s reported later rates are twice its introductory prices

The introductory rates match Argon’s. The Register says Carbon’s prices will then double, although Google has not specified when the introductory period ends. The later rates would match those quoted for Anthropic’s Claude Opus 5.5.

Tokens are the pieces of text used to measure those requests and the model’s replies: input is what the customer sends, and output is what the model generates.

For a customer using the same input and output volumes at these rates, the later prices would double the charge. Google separately says cached input—previously stored input that can be reused—will cost 95% less than the input-token rate.

Google’s tests and external rankings

In the October 28 announcement, Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief SI architect, said Carbon improves secure coding and lengthy engineering tasks. Google’s internal teams have seen Carbon work for days on multistep problems, including finding the causes of subtle bugs and fixing them across large codebases.

Carbon achieved more than twice Argon’s success rate at generating secure code in internal tests. The Chrome team reported that Carbon performed more than ten times better at producing high-quality software fixes than an unnamed rival laboratory’s model. Both comparisons come from Google’s own teams.

Who can apply, and what comes next

Google describes Fairwind as serving trusted, high-priority cybersecurity defenders, including governments, healthcare providers and telecommunications services. Its stated reason for restricting access is that models able to find and fix security flaws could also pose threats in the wrong hands.

Organizations can request access through the application form on the Fairwind Program page. Google checks applicants' security history and record of ethical operations. Participants must use access controls and phishing-resistant multifactor authentication, which requires more than one proof of identity, and cannot share, redistribute or sell model access. Program membership does not promise immediate Carbon access: Google has committed initially to only a subset of defenders.

Google says it is participating in the US government’s voluntary pre-release testing process while collecting feedback and refining safeguards. Those safeguards include defenses against misuse, monitoring of the model’s reasoning, and resistance to prompt injection—malicious instructions intended to redirect the model.

Google has set no release date for ordinary developers, businesses or consumers, saying it wants broader access “as soon as possible.” Its next announced step is access for selected defenders this week, followed by a gradual expansion to additional trusted Fairwind partners.