Google's Gemini 4 Carbon tops the Artificial Analysis index, but only cyber defenders get it for now

Google DeepMind has introduced Gemini 4 Carbon, its new frontier model, only two weeks after shipping Gemini 4 Argon. The announcement, published on October 28, 2026 by Koray Kavukcuoglu, Google's chief AI architect, presents Carbon as the company's strongest model so far, with particular emphasis on cybersecurity and on agentic work that runs over long periods.

A restricted launch

Unlike Argon, Carbon is not broadly available. Google says it is opening access this week to a subset of cyber defenders, then gradually to more trusted partners through its Fairwind Program. Developers, businesses and consumers are promised access "as soon as possible", once Google has collected more feedback and refined its guardrails. The post does not say which products will carry the model, whether the Gemini app, the API or Vertex AI, nor when. Google also says it is taking part in the US government's voluntary pre-release testing of models.

Pricing is already set. The introductory rate matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. It will later rise to $4 and $20. According to The Decoder, the introductory discount runs until the end of 2026.

What the benchmarks show

On the Artificial Analysis Intelligence Index v5.0, Carbon ranks first with 59 points, compared with 50 for Argon. The Decoder notes that the gain is not uniform: Carbon still trails Claude Opus 5.5 on the index's coding-agent measure, and its hallucination rate rose from 4% to 11%, even as answer accuracy improved. On WebDev Arena it climbed 89 points to fourth place.

Google's own security results, which it says were run without guardrails, include 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym and 71.0% on CWE-bench for writing patches. The company also reports a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark and fewer unsupported claims than Argon. Several of these figures come from internal tests that have not been independently checked.

Why access is gated

The same skills that help defenders find and fix flaws could help attackers exploit them, which explains the staged rollout. Google describes internal teams letting Carbon work for days on hard engineering tasks such as tracking down subtle bugs in large codebases. How that holds up outside Google, and how quickly the model reaches ordinary developers, remains to be seen.