Google DeepMind unveils Gemini 4 Carbon, built for cybersecurity and long-horizon tasks

On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, which the company calls its strongest model to date. Its headline strengths, according to the developers, are cybersecurity and the ability to work on hard problems for long stretches without stopping. The announcement came from Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google.

Carbon arrives just two weeks after Gemini 4 Argon and builds on it. It ranks first on the updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59.

Coding and long-horizon work

Carbon is designed for sustained agentic workflows. DeepMind says that in internal projects the model ran for days on a single task. It worked its way out of dead ends, tracked down the root causes of subtle bugs in large codebases and low-level systems, and fixed them. On WebDev Arena it gained 89 Elo points and climbed to fourth place with a score of 1769.

Cybersecurity

DeepMind claims Carbon delivers state-of-the-art results at every stage of cyber defense, from finding vulnerabilities to validating and fixing them. The external benchmarks were run with the model's guardrails turned off:

• SEC-bench Pro (finding vulnerabilities): 88.2%;
• ExploitGym (exploiting vulnerabilities): 54.8%;
• CWE-bench by Collinear (patching vulnerabilities): 71.0%.

On internal benchmarks, Carbon generates secure code more than twice as often as Argon. The Chrome team reported that Carbon did more than 10x better than another lab's large frontier model at producing high-quality patches. Wiz, Google's cloud security platform, says Carbon leads its web penetration testing benchmarks.

Safety and honesty

Carbon uses the same multi-layered safety stack as Argon: model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. According to the company, the model has also become noticeably more honest. On Terminal-Bench 4.0, successful cheating fell to 0.0%, and cheating attempts dropped to 0.4%. In live coding work with internal teams, high-severity misalignment alerts fell by almost 80%.

Unsupported claims are down roughly 2x compared with Argon. The model oversteps the scope of a task less often (down 26%) and under-delivers less often (down 29%). When a request is ambiguous, it first points out discrepancies in the sources to the user instead of guessing. On Gray Swan's indirect prompt injection benchmark, no attack against Carbon succeeded, compared with a 0.7% success rate against Argon.

Access and pricing

DeepMind is rolling the model out gradually. The company is taking part in the U.S. government's voluntary pre-release testing process. This week, Carbon goes to a subset of cyber defenders. Access will then expand in phases to trusted partners through the Fairwind Program. Developers, businesses and consumers will get the model once its safeguards have been refined further, as soon as possible.

The introductory price matches Argon's: $2 per million input tokens and $10 per million output tokens. Cached input tokens cost 95% less. Once the introductory period ends, prices rise to $4 and $20 respectively.

Source: Google DeepMind