Google's Gemini 4 Carbon Tops Intelligence Index at 59, Ships First to Cyber Defenders

Google says Gemini 4 Carbon ranks #1 on Artificial Analysis Intelligence Index v5.0 at 59, but access starts with cyber defenders only.

What happened

Google DeepMind announced Gemini 4 Carbon, which it calls its next frontier model, two weeks after the release of Gemini 4 Argon. Google says Carbon ranks #1 on the recently updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59. The Verge reports that Claude Opus 5.5 and GPT-6 Astra trail behind on that index, and that Google made the announcement on Wednesday.

Google positions the model around cybersecurity defense, reporting state of the art results on three external benchmarks evaluated without guardrails: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on Collinear's CWE-bench for patching. On internal benchmarks, Google says Carbon delivers more than 2x uplift in success rate for generating secure code over Argon, and that the Chrome team reported Carbon performed more than 10X better than another lab's large frontier model at generating high quality patches. Wiz, described by Google as its cloud security platform, observes Carbon leading on web penetration testing benchmarks. On WebDev Arena, Google says Carbon gained 89 elo points to reach #4 at 1768.

On alignment, Google reports successful cheating on Terminal-Bench 4.0 reduced to 0.0% with cheat attempts at 0.4%, high-severity misalignment alerts in internal live coding environments down by almost 80%, unsupported claims cut by approximately 2x versus Argon, overreach reduced 26% and under-service 29%, and a 0.0% attack success rate on Gray Swan's Indirect Prompt Injection Benchmark, where Argon scored 0.7%.

Google says it is engaged in the U.S. government's voluntary process for pre-release model testing. Access starts this week with a subset of cyber defenders, phasing to trusted partners through the Fairwind Program.

Why it matters

For teams that track leaderboards to pick a default model, the reference point has moved: Google claims the top slot on Artificial Analysis Intelligence Index v5.0, with The Verge reporting Claude Opus 5.5 and GPT-6 Astra behind it. Ars Technica adds a caution that matters for procurement decisions: every frontier model scores in the same neighborhood on these tests, and the gap between Carbon and the competition is just a few points. Ars Technica also notes that while Google positions Carbon as a new model, it could be a post-training upgrade of Argon, and that Carbon bests Argon by slim margins in every test Google published.

The claims most relevant to agent builders are the alignment and injection numbers. A reported 0.0% attack success rate on Gray Swan's Indirect Prompt Injection Benchmark, 0.0% successful cheating on Terminal-Bench 4.0, and reduced overreach and under-service all speak to running long, unsupervised tool-using workflows. Google says internal teams have seen Carbon run for days to persist through dead ends on multi-step engineering tasks such as root-causing and patching subtle bugs across large codebases.

None of that is actionable yet for most builders. Carbon is available only to a subset of cyber defenders, expanding through the Fairwind Program, with developer, business and consumer availability described only as coming soon. Pricing is stated as matching Argon's introductory level. Ars Technica's headline frames the release as a frontier model you can't use.

What we don't know yet

When does Carbon reach developers, businesses and consumers, and what are the criteria for joining the Fairwind Program? Is Carbon a distinct model or, as Ars Technica raises, a post-training upgrade of Argon? Google has not published architectural details to settle that. The benchmark results are Google's own reporting and have not been independently reproduced in the sources, including the SEC-bench Pro, ExploitGym and CWE-bench scores, which Google notes were evaluated without guardrails: how do those numbers change with guardrails enabled in a shipping configuration? What is the baseline and methodology behind internal figures such as the 80% drop in high-severity misalignment alerts, the 26% overreach reduction and the 29% under-service reduction? Which "other lab's large frontier model" did the Chrome team compare against for its 10X patch quality claim? And what does the U.S. government's voluntary pre-release testing process cover, or when will its findings be public?

Sources: Google DeepMind / The Verge AI / Ars Technica AI

How this Deep Take is made: it is written with AI assistance from the linked sources, then checked against them sentence by sentence, and every fact links to where it came from.