Google announced Gemini 4 Carbon on October 28, 2026, describing an AI model aimed at cybersecurity defense and sustained coding tasks. The company said initial access would be limited to selected cyber defenders, with broader availability to follow after feedback and further work on safeguards.
Google said Carbon builds on Gemini 4 Argon, which it introduced on October 13. The company reported gains in coding and agentic workflows, alongside improvements in its internal assessments of reliability and model behavior. It said its teams had seen Carbon run for days on engineering tasks, including finding and patching bugs in large codebases.
Google reports internal safety improvements
In internal testing, Google said Carbon generated approximately half as many unsupported claims as Argon. It also reported fewer cheating and shortcut attempts and reduced overreach when completing tasks under project constraints.
For prompt injection, Google said Carbon resisted “every evaluated attack attempt” on Gray Swan’s Indirect Prompt Injection benchmark. That result is limited to the attacks evaluated and does not establish resistance to every possible attack.
The company also said its teams observed lower misalignment and fabrication rates in live coding environments.
A separate evaluation illustrates why the testing context matters. Artificial Analysis measured an 11% hallucination rate for Carbon on AA-Omniscience v1.1, compared with 4% for Argon, while Carbon’s accuracy improved. That external result and Google’s internal reduction in unsupported claims concern different assessments and should not be treated as equivalent measurements.
Carbon leads the updated intelligence index
Artificial Analysis placed Carbon first on version 5 of its Intelligence Index, with a score of 59. Its October 28 evaluation identified strengths in complex reasoning, knowledge and scientific tasks. The leading scores were:
| Model and reasoning setting | Intelligence Index v5 score |
|---|---|
| Gemini 4 Carbon, high | 59 |
| Claude Opus 5.5, max | 58 |
| GPT-6 Astra, max | 57 |
The index update was announced on October 27. It added FrontierCode 2.0 for coding and Terminal-Bench-Science for scientific research workflows, alongside revisions to existing evaluations.
Carbon’s overall lead did not extend to every task. Artificial Analysis placed it second behind Claude Opus 5.5 on its separate Coding Agent Index and said Carbon was not leading in coding and knowledge work across its evaluations.
Restricted rollout and introductory pricing
Google said it was participating in the U.S. government’s voluntary process for pre-release model testing. Its rollout plan called for selected cyber defenders first, followed by additional trusted partners through the Fairwind Program.
The company said it would gather feedback and refine guardrails before expanding access to developers, businesses and consumers. Its announcement did not specify a general-availability date.
Google said introductory pricing would be $2 per million input tokens and $10 per million output tokens, matching Argon’s introductory rates. Cached input tokens would receive a 95% discount, equivalent to $0.10 per million tokens at that introductory input price.
The announcement also specified higher rates after the introductory period: $4 per million input tokens and $20 per million output tokens. Artificial Analysis said the launch discount had no confirmed end date and used standard rates in its pricing comparisons.
