Google launches Gemini 4 Carbon, its top-ranked model, to cybersecurity defenders first

Google DeepMind has released Gemini 4 Carbon, a new AI model the company says is its strongest yet for cybersecurity defense and long, multi-step engineering tasks. The model debuts at the top of the Artificial Analysis Intelligence Index, a benchmark that scores AI systems on reasoning and problem-solving, with a score of 59 on the index’s newly updated version 5.0.

Carbon’s cybersecurity scores stand out. On SEC-bench Pro, a test of finding real software flaws, it reaches 88.2%. On ExploitGym, which checks whether a model can confirm a flaw is actually exploitable, it hits 54.8%. And on CWE-bench, which measures turning a confirmed vulnerability into a working code fix, it scores 71.0%. Google DeepMind says all three are state-of-the-art results, achieved without the guardrails that normally limit the model’s answers. The company also says Carbon more than doubles Argon’s success rate at writing secure code in internal tests, and that Google’s Chrome team found Carbon produced patches more than ten times better than those from an unnamed rival lab’s frontier model.

Rather than open Carbon to the public, Google is starting with a narrow release. This week it is giving access to a subset of cyber defenders, with plans to extend it gradually to more partners through its Fairwind Program. Google says it is also taking part in the U.S. government’s voluntary process for testing models before release. Wider availability to developers, businesses, and consumers is planned “as soon as possible,” though Google has not given a date. Carbon will carry the same introductory price as its predecessor, Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens.

Google also reports a set of internal changes to how Carbon behaves under pressure. The company says attempts to find shortcuts or cheat on the Terminal-Bench 4.0 coding test dropped to 0.4%, with successful cheating falling to zero, while accuracy improved. High-severity misalignment alerts in internal coding environments fell by almost 80%, according to Google, and unsupported claims made by the model dropped by roughly half compared with Argon. On Gray Swan’s Indirect Prompt Injection Benchmark, which tests resistance to hidden instructions smuggled into a model’s input, Google says Carbon blocked every evaluated attack, down from a 0.7% success rate for attackers against Argon.

Beyond security

Carbon’s gains extend past cybersecurity. Google says the model climbed 89 Elo points on WebDev Arena, a ranking of AI models on web development quality, to reach fourth place. Wiz, the cloud security platform Google owns, says Carbon leads its benchmarks for web penetration testing, work it uses in its Scan for Good initiative.

What happens next depends on how the initial group of cyber defenders uses Carbon under Google’s Fairwind Program. Google has not set a date for bringing the model to its consumer and business products, saying only that it will keep gathering feedback on safeguards before that rollout begins.