Google has a new top model, but most people cannot use it yet. Gemini 4 Carbon, announced on October 28 by Koray Kavukcuoglu, a senior vice president at Google DeepMind and Google's chief AI architect, arrives just two weeks after Gemini 4 Argon reached general availability. Google pitches Carbon as its strongest system so far, with the biggest gains in software security work and in tasks that unfold over many hours or days.
The headline number is a score of 59 on version 5.0 of the Artificial Analysis Intelligence Index, which Google says puts Carbon in first place. The Verge, reporting the same day, noted that the index chart places Claude Opus 5.5 from Anthropic and GPT-6 Astra from OpenAI behind it. On security tests run without guardrails, Google reports 88.2% on SEC-bench Pro, which measures discovery of real software flaws, 54.8% on the exploitation benchmark ExploitGym, and 71.0% on Collinear's CWE-bench, which scores working fixes.
Those results explain the unusual rollout. Google says it is taking part in the U.S. government's voluntary pre-release testing process and, from this week, will give Carbon to a limited set of cyber defenders. More trusted partners will be added in stages through what it calls the Fairwind Program. Developers, businesses and consumers will get access through Google's products later, once the company has collected feedback and adjusted safeguards. No date was given.
Pricing is already set. During an introductory period Carbon will cost the same as Argon: $2 for every million tokens sent to the model and $10 for every million it generates, while cached input is charged 95% below the normal input rate. Afterwards the rate doubles to $4 and $20. Google did not say how long the introductory window lasts.
Beyond security, Google describes Carbon as a model for sustained agentic engineering. Internal teams, it says, have watched it keep working for days on hard problems such as tracking a subtle bug to its source in a large codebase and then patching it. On WebDev Arena the model gained 89 Elo points to reach fourth place with 1768. Google also cites in-house evidence: more than twice the success rate of Argon at writing secure code, a report from the Chrome team that Carbon's patches were over ten times better than those of an unnamed rival frontier model, and benchmark leadership in web penetration testing reported by Wiz, which Google calls its own cloud security platform.
A large part of the announcement concerns behavior rather than raw skill. Google says Carbon made no successful cheating attempts on Terminal-Bench 4.0 and tried to cheat in only 0.4% of cases, that high-severity misalignment alerts in live internal coding fell by close to 80%, and that unsupported claims dropped by about half compared with Argon. It reports 26% less overreach and 29% less under-delivery on multi-step plans, and a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark, down from 0.7% for Argon.
Every one of these figures comes from Google, and several rely on internal benchmarks or teams that are part of the company. The cyber numbers were measured with guardrails switched off, which says more about raw capability than about what customers will see. Independent testing will only become possible once access widens beyond the first cohort of defenders, and that timeline remains open.