Google is introducing Gemini 4 Carbon, a new frontier model that scores highly on benchmarks for identifying, exploiting, and patching vulnerabilities. Google will first make the model available to a select group of cybersecurity experts. Developers and companies will have to wait a little longer.
Carbon follows just two weeks after Gemini 4 Argon. On version 5.0 of the Artificial Analysis Intelligence Index, Carbon tops the list with a score of 59. On WebDev Arena, the model rose 89 Elo points, securing fourth place with 1,768 points. Google is primarily targeting long-term agent-based workflows. Internal teams observed Carbon working continuously for days on complex tasks, such as detecting and patching subtle bugs in large codebases and low-level systems.
Detect, validate, remediate
The focus is on security. On SEC-bench Pro, which measures how well a model detects real vulnerabilities, Carbon achieves 88.2 percent. On the ExploitGym exploit benchmark, that figure is 54.8 percent, and on Collinear’s CWE benchmark for patches, it’s 71.0 percent. However, the external tests were conducted without guardrails.
Internally, the numbers are also impressive. According to Google, Carbon generates safe code successfully more than twice as often as Argon. The Chrome team reported that Carbon performed more than ten times better than a large state-of-the-art model from another lab when creating high-quality patches. Google does not specify which lab that is. Additionally, Wiz, Google’s cloud security platform, ranks Carbon at the top of benchmarks for web penetration testing.
Less cheating, more honesty
Google also worked on alignment. On Terminal-Bench 4.0, successful cheating attempts dropped to 0.0 percent, with only 0.4 percent remaining. In live coding environments, serious misalignment warnings decreased by nearly 80 percent. Carbon makes substantiated claims more often. Unsubstantiated claims are roughly halved compared to Argon. Overreach decreased by 26 percent, and under-service by 29 percent. And on Gray Swan’s Indirect Prompt Injection benchmark, not a single attack succeeded, compared to 0.7 percent with Argon.
Google is participating in the U.S. government’s voluntary process to test models prior to release. This week, a select group of cybersecurity professionals will gain access, followed gradually by more trusted partners through the Fairwind Program. Only then will Carbon be made available to developers, businesses, and consumers.
The introductory price matches Argon’s: $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95 percent discount.