Intel / blog

Gemini 4 Carbon is now available

Source: blog.google
Date: Published Nov 10, 2026

Context

Gemini 4 Carbon, the model Google announced on 28 October and first gave only to cyber defenders in its Fairwind Program, is now open to every paying developer on the Gemini API and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google's Antigravity coding tool. It costs the same per token as Gemini 4 Argon, Google's previous top model, released in October. That does not make the two models cost the same to use, and Google's own announcement and migration notes still call Argon the ideal choice for most software engineering and knowledge work.

Google's pricing page lists identical rates for the two models: $2 per million input tokens and $10 per million output tokens through 30 November, then $4 and $20 from 1 December. When Carbon was announced, Google named the later $4 and $20 rates but gave no end date for the introductory half price. The output rate includes thinking tokens, the reasoning a model writes before its answer, and Google's migration notes say Carbon spends more of them per turn than Argon, with higher latency, while Argon gives faster responses and lower token spend per turn. At the same price per token, a turn on Carbon therefore costs more and takes longer than one on Argon. Google publishes no figure for how much more.

Google's guidance on when to pay that extra follows from the design. The announcement keeps Argon as the choice for most software engineering and knowledge work and pitches Carbon at the hardest tasks: complex multi-step coding and debugging, open-ended scientific research and, according to the migration notes, security auditing and dense synthesis, where Carbon uses its extra tokens to explore and check its own work. Moving from Argon needs only a change of model ID. Carbon defaults to a medium thinking level, with low and high available; Google recommends testing all three and using the lowest that meets the quality bar, and raising output limits to at least 32,768 tokens because Carbon can spend more of that limit on thinking. Both models share a one-million-token .

The cost comparisons in the announcement are against other labs' models, not Argon. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at less than half the cost per task of the next best outside model. It also cites second place on the new Multimodal Knowledge Agent benchmark, which Google built with Artificial Analysis, at more than six times lower cost than the leader. Google reports these results and does not name the competing models in this post, so they say where Carbon sits against rivals, not whether a given job justifies Carbon over Argon.

Key takeaways · AI-distilled

Google positions Carbon above Argon for the hardest tasks: it spends more tokens to explore, reason and run verification loops, while Argon stays the recommended default for most coding and knowledge work. Paid API access uses Argon's introductory $2/$10 per million tokens.
Google reports Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at over 2x lower cost per task than the next best external model, and second on the new MKA benchmark at over 6x lower average cost than the leader.
The /swarm preview in Google Antigravity turns one prompt into a tracked team of agents. Google suggests giving Carbon design, coordination and auditing roles while lighter models work in parallel isolated workspaces alongside adversarial verifiers.
Google says Carbon's safeguards cut the per-attempt breach rate in simulated cyberattack testing to 0.8% from Argon's 3.1%, and Carbon passed 92% of a 375-task agentic honesty benchmark versus 70% for Argon.
In a Google setup, Carbon retrieved 40x more remote homologs on average across 7 low-homolog CASP17 protein families, giving structure prediction models more input without retraining. Google notes broader benchmarking is still underway.

Terms in this piece · Glossary

multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.

Read the source blog.google