Google DeepMind made Gemini 4 Carbon generally available on November 10, 2026, to paying developers via the Gemini API and AI Studio, and to Google AI Ultra subscribers. The model had been restricted for approximately two weeks to cybersecurity defenders in Google's Fairwind Program before the broader release. Databricks made Carbon available on its platform on November 11, the same day as general availability, and reported Carbon scored 66.7% on OfficeQA Pro v2, up from 57.8% for Gemini 4 Argon.
Introductory pricing is $2 per million input tokens and $10 per million output tokens through November 30, rising to $4 and $20 on December 1. Carbon supports low, medium, and high thinking effort levels, defaulting to medium. Synchronous requests are capped at 65,536 tokens, and the configured token maximum counts thinking and output together. Third-party sources Artificial Analysis and Vals report a 1,048,576-token context window; Google's own launch post does not state a context window or output limit.
Carbon ranks first on the Artificial Analysis Intelligence Index v5 with a score of 59, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57), leading on four of nine evaluations. It ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, with Google stating per-task cost is more than 2x lower than the next best external model. Cursor confirmed Carbon's availability and reported a top score of 60.4% on CursorBench at high effort. The Neuron Daily notes that Google's top benchmark rankings are on its own or partner-run tests. Carbon does not lead everywhere: on the Coding Agent Index, Carbon with Antigravity CLI scores 59 while Claude Code with Opus 5.5 scores 62. Carbon trails Opus 5.5 on FrontierCode v1.1 (51% vs. 56%) and lags on FrontierSWE 2 and OSWorld. On agentic knowledge work (AA-Briefcase), Carbon's Elo of 1609 falls behind Grok 4.7 (1638), Muse Wonder (1659), and Kimi K3.1 Pro (1678), though up 126 points from Argon.
In cybersecurity, Google reports Carbon scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, all without guardrails. A reviewer identified that the ExploitGym headline figure used a 10-hour compute cap against only 6 hours for Opus 5.5, and that Carbon scores 53.5% under equal conditions. Google states Carbon is the first model to achieve a 0% prompt injection attack success rate on the Gray Swan IPI benchmark. Via the Fairwind Program, Carbon hit 100% recall on scanning and 95% precision on triage in Google's internal Project Fortify benchmark, and blocked 99% of attack trajectories from an adaptive red-teaming agent. Google's Chrome team reports Carbon performed more than 10x better than a competing frontier model on generating high-quality patches, though that comparison names no competing model and provides no supporting numbers.