Google DeepMind unveils Gemini 4 Carbon

Google DeepMind announced Gemini 4 Carbon, its new frontier model, two weeks after it launched Gemini 4 Argon. The company calls Carbon its strongest model so far, with a focus on cybersecurity and long-running agentic work.

Google says Carbon ranks #1 on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. On cybersecurity benchmarks, which Google says it ran without guardrails, Carbon scored 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on Collinear's CWE-bench (patching). Google calls these state-of-the-art results. On WebDev Arena, Carbon gained 89 Elo points to reach #4 at 1769.

Artificial Analysis confirmed the 59 and found Carbon posts the top score in 4 of the Index's 9 evaluations, but does not lead in coding or knowledge work. On the firm's Coding Agent Index, Carbon scored 59 with Antigravity CLI, second only to Claude Code with Claude Opus 5.5 (62), and ranked second on Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%). It led AA-Omniscience v1.1 with 54, though its hallucination rate rose from Argon's 4% to 11%. At standard pricing, an Index task costs $7.76, about 23% less than with Claude Opus 5.5.

Google also cited internal results that have not been independently checked:

• More than double Argon's success rate at generating secure code.
• The Chrome team found Carbon did more than 10x better than "another lab's large frontier model" at writing high-quality patches. Google did not name that model.
• On Terminal-Bench 4.0, successful cheating fell to 0.0% and cheating attempts to 0.4%.
• In live coding with internal teams, high-severity misalignment alerts fell by almost 80%.
• Unsupported claims were roughly half Argon's rate.

Internally, Google says Carbon has run for days on complex engineering tasks, such as tracking down and patching subtle bugs in large codebases.

Access is limited at first. Google says it is taking part in the U.S. government's voluntary pre-release testing process. This week it will open Carbon to a subset of cyber defenders, then add more trusted partners through its Fairwind Program. Google gave no date for access by developers, businesses or consumers. When Carbon does launch, it will cost the same as Argon: Artificial Analysis says that is $4 per million input tokens and $20 per million output tokens at standard rates, with a 50% launch discount and no confirmed end date. Cached input is 95% off.