On October 28, 2026, Google announced Gemini 4 Carbon, which it calls "our strongest model with state-of-the-art cybersecurity and long horizon capabilities". The release is unusual in who gets it first. Google says that "starting this week" it will make Carbon available "to a subset of cyber defenders" and will "gradually phase access to additional trusted partners" through its Fairwind Program. Developers, businesses and consumers come later, with no date given.
Why the rollout is staged
Google says it is "taking a deliberate approach to rollout" and is "actively engaged in the U.S. government's voluntary process for pre-release model testing". It does not report results from that testing. It says it is still iterating on guardrails before a wider release. The order matters: a model that is good at finding and exploiting vulnerabilities is offered first to the people who patch them, while the safeguards for everyone else are still being tuned.
What Google reports on security work
Google reports a lead on SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for exploitation and CWE-bench at 71.0% for patching. It notes that these external benchmarks were "evaluated without guardrails", so they describe raw capability, not what a defender will see through the product. On its own coding tests, Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts at 0.4%, and that high-severity misalignment alerts in live internal coding environments dropped "by almost 80%". These are internal measurements.
What the independent index shows
Artificial Analysis, also on October 28, puts Carbon first on its Intelligence Index v5 at 59, with the top score in four of nine evaluations. It is more mixed on coding. With Antigravity CLI, Carbon scores 59 on the Coding Agent Index, three points behind Claude Code with Claude Opus 5.5 (max) at 62. It ranks second on Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%, against Opus 5.5's 56%).
One number moved the wrong way. On AA-Omniscience, Carbon's hallucination rate rose from Gemini 4 Argon's 4% to 11%. Artificial Analysis notes that this is still the second lowest of any model it tracks, against 50% for both Claude Opus 5.5 (max) and GPT-6 Astra (max).
What it costs to run
Google lists introductory prices of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. Artificial Analysis uses the standard prices and reports $7.76 per Intelligence Index task, against $10.04 for Claude Opus 5.5 (max) and $6.25 for GPT-6 Astra (max), with Carbon using 123k output tokens per task against Opus 5.5's 173k. Cost per task, not price per token, is the figure to compare.
What remains uncertain. The cyber benchmarks were run without guardrails, and the shipped product will have them. The safety and honesty gains are Google's internal tests. Google has not said when general access starts or when the introductory price ends, and the outcome of government pre-release testing is not public.
Analysis: For engineering teams the practical point is that a frontier model can now arrive in stages, with defenders first. If a team at Belgavi's Book Store wanted Carbon to review the code that handles order refunds, it would have to wait for general access or qualify as a defender. In the meantime, the useful work is to prepare an evaluation set from its own past security bugs, so the model can be compared on the store's code once access opens. The rise in hallucination rate is also a reminder that a model can lead an index and still get worse on one property a team depends on. Check the metric your workflow needs, not the headline rank.
Sources and evidence
Google announced Gemini 4 Carbon on October 28 and is giving it first to a subset of cyber defenders through its Fairwind Program, with developer access later and no date. Artificial Analysis ranks it first on Intelligence Index v5 at 59 but second on coding agents (59 against Claude Code with Opus 5.5 at 62), and its hallucination rate rose from Argon's 4% to 11%; Google's cyber scores were run without guardrails.