Google has introduced Gemini 4 Carbon, its latest AI model, which ranks first in the Artificial Analysis Intelligence Index V5 with a score of 59. It is one point ahead of Claude Opus 5.5 and two points ahead of GPT-6 Astra, while posting the highest score in four of the index’s nine evaluations.
The company is positioning Gemini 4 Carbon for coding, cybersecurity and complex task execution. Google says the model can work autonomously on complex software-engineering projects for multiple days, identifying bugs and proposing fixes.
- SEC-bench Pro: 88.2% for identifying software vulnerabilities.
- ExploitGym: 54.8% for evaluating vulnerability exploitation.
- CWE-bench: 71% for assessing its ability to remediate software issues.
Google also reports improvements in factual reliability. In the AA-Omniscience v1.1 evaluation, which measures information accuracy and the propensity to generate unsupported answers, the company says Gemini 4 Carbon cut unsupported claims by roughly half compared with Gemini 4 Argon, released two weeks earlier. It also recorded a 0% success rate for prompt-injection attacks in the Gray Swan evaluation.
Those benchmark results do not make the model infallible. Gemini 4 Carbon still trails some competitors in certain coding and reasoning tests.
Google has set introductory pricing at $2 per million input tokens and $10 per million generated tokens, with a 95% discount for cached input tokens. Access is currently restricted to cybersecurity specialists and selected partners through the Fairwind program. Google plans a gradual expansion to developers, businesses and general users, but has not announced a timetable.
