Gemini 4 Carbon scored 88.2% at finding software flaws and opened first to cyber defenders
Google DeepMind says the model also reached 54.8% on exploitation tests and 71.0% on patching, at $2 per million input tokens.
In brief
Google's new model scored 88.2% on a flaw-finding test
Gemini 4 Carbon leads three security benchmarks, by Google's own account. It goes first to a small group of cyber defenders. Public release has no date yet.
Google DeepMind has announced Gemini 4 Carbon, a model it calls its strongest yet, with a focus on cybersecurity and long, multi-step engineering work. The claims come from Google's own launch post, not from independent testing.
Who gets it first
Carbon is not open to everyone. This week it goes to a small group of cyber defenders. More trusted partners will follow through the Fairwind Program. Google says it is also taking part in the U.S. government's voluntary process for testing models before release.
Developers, businesses and consumers come later, with no date given beyond "as soon as possible".
The security numbers
Google reports three scores on outside benchmarks, tested without guardrails:
- SEC-bench Pro, finding vulnerabilities: 88.2%
- ExploitGym, checking whether a flaw can be exploited: 54.8%
- CWE-bench, fixing the root cause: 71.0%
Google says all three are state of the art. On its internal tests, Carbon more than doubled Argon's success rate at writing secure code. Argon is the model launched two weeks earlier.
The Chrome team reportedly found Carbon more than 10 times better than another lab's frontier model at producing good patches.
Long jobs and general ranking
Google says teams saw Carbon work for days on hard engineering problems, such as tracking down subtle bugs in large codebases. It also ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. On WebDev Arena it rose 89 Elo points to 1768, placing fourth.
Behaving better
Google reports fewer shortcuts and fewer made-up claims. On Terminal-Bench 4.0, successful cheating fell to 0.0% and cheating attempts to 0.4%. Unsupported claims dropped by about half against Argon. High-severity misalignment alerts in live internal coding fell by almost 80%. On Gray Swan's prompt injection benchmark, which tests whether hidden instructions can hijack a model, Carbon had a 0.0% attack success rate, against 0.7% for Argon.
Pricing is the same as Argon's: $2 for each million input tokens and $10 for each million output tokens. Cached input costs 95% less.
Imagine a security team that must fix thousands of flaws. Which ones are real, and which come first? That is the job Google is aiming at.
The limits
Most figures are Google's own or from internal tests. The source does not say when the public can use Carbon, and it gives no independent verification. Google says it is still collecting feedback and adjusting guardrails before wider release.
Source: Google DeepMind Blog