Google has announced Gemini 4 Carbon, positioning it as a frontier model for complex coding and cybersecurity work. The company says it is stronger on long-horizon agentic tasks, aiming to sustain multi-step engineering work rather than only answer one-off prompts.
Gemini 4 Carbon’s benchmarks and engineering focus
Google reports a score of 59 on Artificial Analysis Intelligence Index v5.0, alongside results of 88.2%, 54.8%, and 71.0% on benchmarks for finding vulnerabilities, validating exploits, and fixing flaws. These figures reflect specific tests and do not establish performance across all real-world environments. The company also says internal testing found more than a twofold increase over Argon in success at generating secure code; those internal results are not independent validation.
Gemini 4 Carbon will roll out gradually, with safety still to assess
The announcement also emphasizes model behavior. Google says internal tests found fewer unsupported claims and a lower success rate for prompt-injection attacks. It says Carbon is designed to retrieve sources and flag discrepancies when a task is ambiguous instead of guessing. These are company-reported evaluations; users and independent testing will still need to establish how reliable the behavior is in practice.
Google plans to start with selected cyber defenders, expand access to trusted partners, and later offer Carbon to developers, businesses, and consumers as safeguards are refined. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens; after the introductory period, those prices rise to $4 and $20. The staged rollout reflects the implications of dual-use capabilities: a model that can help find and fix vulnerabilities may also require careful access controls. The timing of broader availability, independent evaluation, and performance at scale remain open questions.