If you plan to build on Gemini 4 Carbon, the date to note is 30 November. Google has released Carbon, which it calls its strongest model, to paying developers through AI Studio and to Google AI Ultra subscribers in the Gemini app and Antigravity, at $2 per million input tokens and $10 per million output. That is an introductory price; afterwards it doubles to $4 and $20. A workload that looks affordable this month will cost twice as much in December.
The cost picture is more complicated than the rate card. Google says Carbon deliberately uses more tokens than Gemini 4 Argon on each task, exploring and verifying more thoroughly, and recommends Argon for most everyday engineering and knowledge work. At the same time it claims Carbon is cheaper per completed task than rival frontier models, more than 2X lower than the next best external model on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. Both things can be true; the only way to know which dominates for your work is to measure tokens per job.
The safety figures are the other notable part of this release. On Google's 375-task agentic honesty benchmark, Carbon passed 92% of tasks against 70% for Argon, declining to invent missing experimental data 98% of the time and flagging methodological flaws 68% of the time. In simulated cyberattack tests, Google says Carbon's safeguards blocked 99% of attack attempts from an adaptive red-teaming agent and cut the per-attempt breach rate to 0.8%, down from 3.1% on Argon. External testers at Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests.
There is also a research angle. Google is extending its programme that offers 10,000 scientists free access to its top Antigravity tier for six months to include Carbon, and cites work such as using Carbon to find 40-fold more distant protein relatives on average across 7 hard protein families, and the AlphaProof Code verification of Signal. For readers comparing Carbon with Claude Opus or GPT-class models, the honesty results are worth weighing alongside raw benchmark scores, because agents that fill in missing data are a real failure mode.
The caveats are the usual ones for a vendor launch. The honesty benchmark and most safety numbers are Google's own. Google itself says prompt-injection robustness still needs work. And the comparisons with other labs' models depend on test set-ups that outsiders have not yet reproduced.