Google releases Gemini 4 Carbon to paying developers at $2 per million input tokens
It leads LMArena's text vote, trails Claude Fable 5.2 on Artificial Analysis (60.0 vs 64.6) and has no Epoch AI overall score yet.
Google made Gemini 4 Carbon, its strongest model, available on 10 November 2026. Paying developers and enterprises can use it through the Gemini API, and Google AI Ultra subscribers get it in the Gemini app. Until now only cyber defenders in Google’s Fairwind Program had access. The price is $2 per million input tokens and $10 per million output tokens (a token is a short piece of a word). That matches Gemini 4 Argon, and Google’s changelog says it is an introductory price through 30 November.
Independent scoreboards place it near the top but not first. On the Artificial Analysis Intelligence Index, which combines about ten tests, it scored 60.0. Claude Fable 5.2 scored 64.6 and GPT-6 Astra 56.9. On LMArena’s text leaderboard, where people vote blind between two answers, it leads with 1526.6, ahead of Gemini 4 Argon (1522.0) and within the margin of error. Claude Fable 5.2 is not listed there yet.
On long software jobs it sits between the two rivals. FrontierSWE, 34 engineering tasks with up to 20 hours each, gave Carbon 61.2, against 69.9 for Fable 5.2 and 65.5 for GPT-6 Astra. On APEX-Agents, office tasks written by lawyers and bankers, Carbon scored 84.2%, ahead of Fable 5.2 (80.1%) and GPT-6 Astra (64.7%).
Google also reports its own results. It says Carbon passed 92% of the tasks in a 375-task agentic honesty test, against 70% for Argon. It also says Carbon ranks first on CursorBench 4.0 at more than 2× lower cost per task than the next external model.
What it means
Google’s best model is now open to anyone paying for the API, at the price of its second-best. Carbon is ahead of GPT-6 Astra on two of the three independent task tests above and behind Fable 5.2 on two. Nobody outside Google can yet say how it ranks overall on Epoch AI’s index.
Caveats
- Epoch AI’s Capabilities Index has no row for Carbon as of 10 November, so it has no overall rank on the site’s main measure.
- The honesty result and the CursorBench ranking are Google’s own claims. The post does not link the underlying runs.
- The $2 and $10 price is introductory and ends on 30 November. The post says Carbon uses more tokens than Argon on hard tasks, so cost per task can differ from the list price.
- The LMArena figure is from 9 November, before the public launch, and its margin of error overlaps Argon’s score.
- Scores for Carbon were run on different settings: Artificial Analysis at “high” effort, FrontierSWE with Proximal’s harness, and APEX-Agents at an unknown effort.