Google releases Gemini 4 Carbon to cyber defenders a month after Argon
Artificial Analysis ranked it first on its new v5.0 index at 59, a point above Claude Opus 5.5; introductory pricing matches Argon at $2/$10 per million tokens.
- Models & capabilities
- Security & misuse
- Major
Google announced Gemini 4 Carbon, which it called its strongest model, four weeks after Gemini 4 Argon’s restricted launch and two weeks after Argon’s general release. Gemini 4 Carbon follows the same staged path. A subset of cyber defenders in Google’s Fairwind programme get it first, then other trusted partners, before developers, businesses and consumers “as soon as possible”, with no date given. Google said it was again taking part in the US government’s voluntary pre-release testing. The introductory price matches Argon’s, $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20. The post was signed by chief AI architect Koray Kavukcuoglu.
Google pitched Carbon at long-horizon engineering and cyber defence. It said internal teams had seen the model work for days on tasks such as root-causing bugs across large codebases. On security benchmarks run “without guardrails”, it reported 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching them. The methodology document shows the first two were Google’s own runs, SEC-bench Pro on a 183-task JavaScript-engine subset. Google also said its Chrome team found Carbon more than ten times better at writing patches than an unnamed rival lab’s model.
Independent leaderboards supported part of the claim. Artificial Analysis placed Carbon first on version 5.0 of its index, released this month, at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. It also topped Zapier’s AutomationBench board at 53.9% and Andon Labs’ Vending-Bench 2. It did not lead everywhere. Google’s own table had Astra ahead on FrontierSWE 2 and Opus 5.5 ahead on OSWorld 2.1, and on Artificial Analysis’s Terminal-Bench 4.0 runs Kimi K3.1 Pro scored higher.
On safety, Google said Carbon used Argon’s safeguards and had been trained further against shortcuts. It reported successful cheating on Terminal-Bench 4.0 down to zero, with attempts at 0.4%, high-severity misalignment alerts in internal coding work down by almost 80%, and unsupported claims roughly halved against Argon. All of these were Google’s internal measurements. The launch post pointed to Argon’s safety report rather than a new one.