AI News
Google Gemini 4 Carbon: Its Strongest Model for Cybersecurity, and Almost Nobody Can Use It Yet
Google announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon. Google calls it its strongest model yet, built for two jobs: long, multi-step engineering work and cybersecurity defense — finding vulnerabilities, proving they are exploitable, and patching them at the root cause.
What Google claims
Carbon ranks #1 on the updated v5.0 Artificial Analysis Intelligence Index at 59, ahead of Claude Opus 5.5 (58), GPT-6 Astra (57) and Argon (50). In cybersecurity tests it leads: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploitation, 71.0% on CWE-bench for patching. In web development it jumped 89 Elo points to #4 (1768) on WebDev Arena. Google says its internal teams have watched Carbon run for days on single engineering problems, persisting through dead ends to root-cause subtle bugs across large codebases.
The safety claims are just as strong: secure-code success more than doubled over Argon; high-severity misalignment alerts fell almost 80% in live internal coding; unsupported claims were cut roughly in half; overreach fell 26% and under-service 29%; on Gray Swan’s indirect prompt-injection benchmark, Carbon resisted every attack (0.0% success, against Argon’s 0.7%); cheating on Terminal-Bench 4.0 fell to 0.0%.
Who can use it — almost nobody, yet
Starting this week, Carbon goes to a subset of cyber defenders, then gradually to additional trusted partners through Google’s Fairwind Program. Developers, businesses and consumers get it later, after Google gathers defender feedback and iterates on safeguards. Google says it is also participating in the US government’s voluntary pre-release testing process. There is no public signup page and no date for wider access. Pricing is set at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off.
What to believe and what to check
The Artificial Analysis index is independent, but the version it ranks on (v5.0) was just updated, and Carbon’s lead over Opus 5.5 is one point. Most other headline numbers are Google’s own, from internal benchmarks — including the 10x patch claim, which compares against an unnamed rival lab’s model. Google itself notes the external cyber benchmarks were evaluated without guardrails. Independent coverage notes rivals still lead on some tests, and one analysis found a higher wrong-answer rate than Argon on one test. The honest position: Carbon looks like a genuine step for defensive security work, but it is not yet a model anyone outside a vetted group can test. And any site claiming to sell Carbon access today is not selling Google’s model.