Google previews Gemini 4 Carbon to cyber defenders first, $2/$10 per 1M tokens

Google announced Gemini 4 Carbon on 28 October 2026, calling it its strongest model for cybersecurity and long-horizon work. It is not generally available. Google says it will give access to a subset of cyber defenders that week and widen it gradually to trusted partners through its Fairwind Program. It gives no date for developers, businesses or consumers, saying only "as soon as possible".

What Google reports

• Cybersecurity: 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on Collinear's CWE-bench (patching). Google says these were run without guardrails.
• Coding: Google reports that the Chrome team saw Carbon perform more than 10 times better than another lab's frontier model at generating high-quality patches. Google names neither the model nor the method.
• Behaviour: Google says successful cheating on Terminal-Bench 4.0 fell to 0.0%, high-severity misalignment alerts in internal live coding dropped by almost 80%, unsupported claims roughly halved against Argon, and the Gray Swan indirect prompt-injection attack success rate went from 0.7% to 0.0%.
• Long tasks: Google says internal teams saw Carbon run for days on multi-step engineering tasks such as root-causing bugs in low-level systems.

Every figure in that list is Google's, from its own post, and the behaviour numbers come from internal testing. Google also says it is taking part in the US government's voluntary process for pre-release model testing.

What it costs

Google says Carbon will launch at the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper. A footnote says that after the introductory period the price becomes $4 and $20. The post does not say when that period ends. Artificial Analysis describes the same arrangement as a 50% launch discount with no confirmed end date, and uses the $4 and $20 standard price in its charts.

What an independent test found

Artificial Analysis reports that Carbon (high reasoning) scores 59 on its Intelligence Index v5, first overall. On its Coding Agent Index it scores 59 with Antigravity CLI, 3 points behind Claude Code with Claude Opus 5.5 at 62. It puts the cost at $7.76 per Index task at standard prices, about 23% less than Opus 5.5 at $10.04, and says Carbon used 123,000 output tokens per task against 173,000.

It also found a cost on honesty. On its AA-Omniscience v1.1 test Carbon scores 54, ahead of Opus 5.5 at 48, but its hallucination rate rose from 4% for Argon to 11%. Artificial Analysis notes that this is still the second lowest of the models it tested. On its AA-Briefcase knowledge-work test Carbon scored 1609 Elo, up 126 on Argon but behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro.

What it means for a rollout

If your team is not a security vendor or an approved defender, nothing changes today: Argon remains the Gemini model you can call. If you plan security-review or patching agents, Carbon is the model to put on a watchlist, and the 11% hallucination figure is a reason to keep a human check on any vulnerability report it writes. Budget at the $4 and $20 rate, not the introductory one, since the end date is unknown.