Gemini 4 Carbon is now available

Context

Gemini 4 Carbon, which Google DeepMind calls its strongest model, is now open to paying developers and businesses, two weeks after Google gave it first to cybersecurity defenders in its Fairwind Program (announcement). Developers reach it through the paid tier of the Gemini API in AI Studio, and businesses through Gemini for business. Google AI Ultra subscribers also get it in Google Antigravity, Google's agentic coding environment, the Gemini app and AI Mode in Search. The version tuned for security work, Carbon for cyber, still requires an application to Fairwind.

The practical question is when to use Carbon instead of Gemini 4 Argon, released a few weeks earlier. Google's answer is that Argon remains the right choice for most software engineering and knowledge work, while Carbon is meant for problems at the limit of a model's ability, such as complex multi-step debugging or open-ended scientific research. Carbon gets there by spending more than Argon to explore, reason and run repeated verification loops. Both models carry the same introductory per-token price, so a task Carbon works through with more tokens costs more than the same task on Argon. Google's cost claim is against other companies' models: it says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next-best external model. The post gives the rankings without scores.

The launch price is temporary. Carbon costs $2 per million input tokens and $10 per million output tokens, the same introductory price as Argon, but a footnote to Google's post says that after the introductory period ends on November 30, 2026, the price becomes $4 and $20. Anyone measuring cost per task in the first weeks should therefore expect it to double for the same token usage. Google's earlier Carbon announcement covers the security benchmarks and the close Artificial Analysis index ranking behind the "strongest model" claim.

Key takeaways · AI-distilled

Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best external model, and second on the new Knowledge at over 6x lower cost than the leader.
Antigravity's previewed /swarm feature turns one prompt into a coordinated team of agents, so developers can give Carbon the architecture, coordination and auditing roles while faster models work in parallel isolated workspaces checked by adversarial verifiers.
Google reports Carbon's safeguards blocked 99% of an adaptive red-team agent's simulated attack trajectories, cutting the per-attempt breach rate to 0.8% from Argon's 3.1%, and that it passed 92% of a 375-task agentic honesty benchmark versus 70% for Argon.
In Google's Project Fortify, Carbon reached 100% recall on scanning and 95% precision on triage on an internal benchmark, and on open-source repositories it showed more than a 25% relative recall improvement over Argon.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
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.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.