Google's Gemini 4 Carbon tops Artificial Analysis index with staged defender-first rollout

Google announced Gemini 4 Carbon on October 28, 2026 — its second frontier model inside a single month — and the model has immediately taken first place on the independently run Artificial Analysis Intelligence Index v5.0 with a composite score of 59, according to a write-up on dev.to. The lead over the previous front-runner is a single point, but the detail with broader consequences is access: the rollout starts with cybersecurity defenders, and general availability has no date.

A one-point lead on the composite index

The composite standings reported by the write-up put Gemini 4 Carbon (high setting) at 59, Claude Opus 5.5 (max) at 58, GPT-6 Astra (max) at 57 and Gemini 4 Argon at 50. That is a nine-point jump over Argon, which launched two weeks earlier, and a one-point edge over Opus 5.5.

The composite ranking hides task-level losses, though. On the AA Coding Agent Index, Carbon scores 59 against 62 for Opus 5.5, and on AA-Briefcase it trails Kimi K3.1 Pro, Muse Wonder and Grok 4.7. The best composite score does not mean the best model for every job.

Positioned around cybersecurity

The launch is framed around security. Google describes Carbon as its strongest model to date, with cybersecurity and long-horizon capabilities as the headline claims, and internal teams are said to have run it for days at a time on multi-step engineering tasks.

The reported security results were produced with guardrails removed: 88.2 percent on SEC-bench Pro at finding vulnerabilities, 54.8 percent on ExploitGym at exploiting them, and 71.0 percent on CWE-bench at patching flaws to the root cause. Two caveats apply. These are Google's own figures and have not been reproduced independently, a point the write-up attributes to reporting by The Decoder, and defenders will actually use a safeguarded version that differs from the system benchmarked. On evaluation integrity, Google says successful cheating on Terminal-Bench 4.0 fell to 0.0 percent, with only 0.4 percent of attempts even tried.

Priced, but not usable

Pricing is published while the API is not. Introductory rates are 2 and 10 dollars per million tokens, rising to 4 and 20 after an unspecified period, with cached input discounted by 95 percent. On cost per task, Artificial Analysis puts Carbon at 7.76 dollars — marginally above Argon at 7.68, above GPT-6 Astra (max) at 6.25 and below Opus 5.5 (max) at 10.04.

Access is staged: a subset of cyber defenders receives the model first, then vetted partners through what Google calls the Fairwind Program, then everyone else as soon as possible, with no date given. There is no public API model identifier, no context-window specification and no model card. Ars Technica's summary of the practical situation, as cited in the write-up, is blunt: the model cannot be used.

Why it matters

For builders, the practical advice in the write-up is to keep building on Argon, which is available today and marginally cheaper per task. The deeper signal is the release pattern itself. Frontier capability is now shipping with staged, defender-first access — a pattern the write-up says also involves government pre-release review — and it argues this launch looks like a template rather than an exception. Teams planning products around frontier models should design for staged access from the start. One caution applies to readers as well: everything here rests on a single condensed write-up, and the self-reported security benchmarks in particular await independent confirmation.