Google DeepMind has released Gemini 4 Carbon, the latest iteration in its frontier SI model lineup, positioning it as a specialized tool for high-stakes engineering and cybersecurity defense. The launch highlights a strategic shift toward models optimized for long-horizon agentic tasks and secure code generation, distinct from general-purpose assistants.

What Happened

Gemini 4 Carbon is described by the company as its strongest SI model to date, with specific improvements in cybersecurity and long-horizon capabilities. According to Google DeepMind, Carbon ranks #1 on the recently updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59. The model demonstrates significant gains in coding and agentic tasks, particularly in cybersecurity defense. It achieves 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Internally, the company reports that Carbon shows reduced misalignment and fabrication rates in live coding environments.

The model is designed to sustain complex, multi-step engineering workflows, such as root-causing and patching subtle bugs across large codebases. Google DeepMind states that internal teams have observed Carbon running for days to persist through dead ends. In web development, the model jumps 89 elo points to reach #4 with a score of 1768 on WebDev Arena. Carbon builds upon Gemini 4 Argon, which was launched two weeks prior, illustrating the rapid pace of capability development within the lab.

Rollout will be deliberate. Starting this week, access is limited to a subset of cyber defenders and trusted partners through the Fairwind Program. The company is engaged in the U.S. government’s voluntary process for pre-release model testing. Pricing matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.

Why It Matters

For developers and security professionals, Gemini 4 Carbon represents a targeted advancement in SI safety and reliability for critical infrastructure tasks. The model’s performance on external benchmarks, evaluated without guardrails, suggests a leap in autonomous vulnerability management. Google DeepMind reports that Carbon achieves more than a 2x uplift in success rate for generating secure code compared to Argon. Additionally, the Chrome team reportedly found that Carbon performed more than 10X better than another lab's large frontier model when generating high-quality patches.

The release also emphasizes alignment and safety improvements. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts dropping to 0.4%. High-severity misalignment alerts in internal live coding environments dropped by nearly 80%. The model also shows improved resistance to prompt injection, achieving a 0.0% attack success rate on the Gray Swan’s Indirect Prompt Injection (IPI) Benchmark, up from Argon’s 0.7%. These metrics indicate a focus on reducing fabrication and overreach, with Carbon reportedly reducing overreach by 26% and under-service by 29%.

The Bottom Line

Google DeepMind is expanding its SI model offerings with a version specifically tuned for cybersecurity and long-horizon coding tasks. While not yet available to the general public, the model’s benchmark results and safety metrics position it as a specialized tool for enterprise and security applications. The company plans to gradually expand access to developers, businesses, and consumers after gathering feedback from the initial cohort.

Sources: Google DeepMind