Google announces Gemini 4 Carbon, #1 on the Artificial Analysis Intelligence Index v5 (59), two weeks after Argon; first only for Fairwind cyber defenders

On Oct 28, 2026, Google announced Gemini 4 Carbon, the second model of the Gemini 4 family, fifteen days after Gemini 4 Argon became generally available. Sundar Pichai, Demis Hassabis, Koray Kavukcuoglu and Logan Kilpatrick posted within two minutes of each other; Pichai’s post had about 677,000 views on the same day. Kilpatrick: “it’s a big jump over Argon which we released just 2 weeks ago. Gemini 4 Carbon is SOTA on cyber, long horizon tasks, ranks #1 on Artificial Analysis V5, significantly improves on usability in domains like coding, and is more aligned.”

The rollout copies Argon’s: a subset of cyber defenders gets it “starting this week”, then more partners in the Fairwind Program, then developers, businesses and consumers “soon”. Google says it is taking part in the US government’s voluntary pre-release testing. The announced price is the same as Argon’s ($2 / $10 introductory, then $4 / $20).

Google reclaimed the top independent overall score two weeks after Argon trailed Opus 5.5 and GPT-6 Astra, and kept the pattern of releasing its strongest models to cyber defenders first. The gap at the top is now one index point, and on coding agents Anthropic still leads.

• Artificial Analysis (posted 21 minutes after the launch) puts Carbon (high) first on its Intelligence Index v5 at 59, one point ahead of Claude Opus 5.5 at max effort and two ahead of GPT-6 Astra at max. Carbon is second, not first, on AA’s Coding Agent Index (59 vs 62 for Claude Code with Opus 5.5) and is not leading on agentic knowledge work (AA-Briefcase). Its hallucination rate on AA-Omniscience rose from 4% (Argon) to 11%, still far below Opus 5.5 and Astra (50%).
• Vals AI puts Carbon second on the Vals Index (70.36%), behind Claude Sonnet 5.5 (70.85%) and level with Opus 5.5 within error.
• Google’s own table shows Carbon behind GPT-6 Astra on FrontierSWE 2 and behind Opus 5.5 on OSWorld-2.1; most rival figures in it are the rivals’ self-reported numbers.

Intelligence Index v5 is a re-weighted version of AA’s index; scores are not comparable with v4 numbers quoted elsewhere in this dataset (for example Argon scored 53 on the earlier version and 50 on v5).

Google claims Carbon is “more aligned” than Argon (fewer cheating attempts on Terminal-Bench, ~80% fewer high-severity misalignment alerts in internal coding, 0.0% indirect prompt-injection success). These are internal measurements. Unlike Argon’s broad release on Oct 13, the Carbon launch came without a model card or a Frontier Safety Framework statement on critical capability levels: the post links the Oct 13 Gemini 4 Safety and Security Report, which predates Carbon. For Argon, Google said it could not rule out the cyber and biological CCLs; whether Carbon was evaluated against them is not yet public. The cyber scores Google highlights include exploitation (ExploitGym 54.8%, run “with mitigations off”), which is why access starts with vetted defenders.

• No public API id, docs page, context window or knowledge cutoff from Google yet (AA lists a 1M-token context).
• No Carbon model card or FSF results; check when broader availability comes.
• The Chrome “10X better” patch comparison names no rival model.
• Arena’s own post (Oct 28) confirms WebDev #4 at 1768 and adds #1 in Text Arena and #2 in Agent Arena; the arena.ai pages themselves were not read.