Google has released Gemini 4 Carbon, its newest AI model, to a limited group of cybersecurity experts before any wider launch, the company said Wednesday.
Carbon arrives about two weeks after Gemini 4 Argon. Google says it is better at finding and fixing software flaws, and at long engineering jobs that can run for days.
Google is working with the U.S. government on voluntary testing before release. It plans to add trusted partners through its Fairwind Program, then open the model to developers, businesses and consumers through its products. No dates were given.
Price
Carbon costs the same as Argon: $2 per million input tokens and $10 per million output tokens, or about Rs 191 and Rs 953. Cached input tokens are 95% cheaper.
How it scored
Google reported these results:
- SEC-bench Pro: 88.2% for Carbon, a test of finding vulnerabilities. OpenAI’s GPT-6 Astra and Gemini 4 Argon scored about 85% and 83%.
- ExploitGym: 54.8%, a test of confirming that a flaw can be exploited. GPT-6 Astra scored 42.4%.
- CWE-bench: 71.0%, a test of fixing the root cause of a flaw. Carbon beat Grok 4.7, Argon and GPT-6 Astra.
- Artificial Analysis Intelligence Index v5.0: first place, with 59.
Google’s own testing showed Carbon more than doubling Argon’s success rate at writing secure code. The company said its Chrome team found Carbon more than 10 times better than a rival model at producing quality security patches.
Wiz, Google’s cloud security platform, said Carbon beat other AI models on web penetration testing benchmarks.
Safety
Google said Carbon makes about half as many unsupported claims as Argon. In a test of prompt injection, where hidden text tries to take control of an AI model, no attacks succeeded against Carbon. Argon was breached 0.7% of the time, according to Gray Swan’s benchmark.
