Back to feed
Artificial intelligenceSingle source

Google unveils Gemini 4 Carbon for finding software flaws and building fixes

Google has introduced Gemini 4 Carbon, a model designed to find software vulnerabilities and develop fixes. The company reported scores of 88.2% on SEC-bench Pro, 71% on CWE-bench and 54.8% on ExploitGym.

Preferred on Google
Text size

Image: itc.ua · Author: \u041e\u043b\u0435\u043a\u0441\u0430\u043d\u0434\u0440 \u0424\u0435\u0434\u043e\u0442\u043a\u0456\u043d · source articleEditorial excerpt for news reporting

Google has introduced Gemini 4 Carbon, a model for finding software vulnerabilities, assessing their severity and developing fixes. The company also describes it as a system for complex engineering work involving extended, multi-step processes and multiple tools.

Carbon scored 88.2% on SEC-bench Pro, which tests vulnerability detection in real-world software. It scored 71% on CWE-bench, which evaluates vulnerability fixes, and 54.8% on ExploitGym, which uses controlled exploitation tests to assess vulnerabilities.

Google says Carbon outperformed competing models on all three benchmarks. The company introduced it two weeks after Gemini 4 Argon and said the new model also improves at coding, reasoning and autonomous task execution.

Google’s internal test results

According to Google’s internal evaluations, Carbon performed more than twice as well as Argon in secure-code generation. The company’s Chrome team found that Carbon was 10 times better than a competing model at producing high-quality security updates.

Carbon gained 89 Elo points in WebDev Arena, reaching 1,768 points and fourth place on the leaderboard. Google also reported improvements in scientific tasks, multimodal understanding and full-stack web development.

Google said it added safeguards for extended work sessions. The company reported that unsupported claims fell by about half compared with Argon, while critical mismatch alerts in coding work declined by nearly 80%.

In the Gray Swan test, Google reported a 0% attack success rate for Carbon, compared with 0.7% for Argon.

Access and pricing

Google will initially restrict Carbon to a selected group of cybersecurity professionals. Additional partners will be able to access it through the Fairwind program.

Introductory prices match Argon’s: $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount.

What we know

  • Google introduced Gemini 4 Carbon to detect software vulnerabilities and develop fixes.
  • The model scored 88.2% on SEC-bench Pro, 71% on CWE-bench and 54.8% on ExploitGym.
  • Carbon placed fourth in WebDev Arena with 1,768 Elo points.
  • Initial access will go to selected cybersecurity professionals, with pricing matching Argon.
  • Google reported a 0% attack success rate for Carbon in the Gray Swan test, versus 0.7% for Argon.

Trust: Single source

  • The story relies on one publisher. A second independent confirmation has not been established from the cited sources.
  • Sources: ITC.ua.
Related news
View sources1

COMMUNITY

Discussion

—

Loading comments…