Summary
Google says Gemini 4 Carbon is now available through its developer and enterprise products, and is rolling out to Google AI Ultra subscribers. The company positions it for complex, multi-step work and lists API prices of $2 per million input tokens and $10 per million output tokens.
Google says Gemini 4 Carbon is now available to developers, enterprises, scientists and consumers through different access routes. The company describes it as a model for demanding, multi-step work, including complex software engineering, knowledge work and scientific reasoning. It is priced at $2 per million input tokens and $10 per million output tokens on the paid API tier, the same introductory rates Google lists for Gemini 4 Argon.
Access depends on the user and product
Developers and enterprises can use Carbon through the Gemini API in Google AI Studio; Google says it is available to paying customers. The company is also rolling it out to Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. Researchers can apply for access through a programme that offers selected participants six months of use, while cybersecurity defenders can apply through Google’s Fairwind Program.
Google is also previewing /swarm in Antigravity. The feature is designed to coordinate multiple agents on a project: a user can assign Carbon to work that calls for deeper reasoning, such as architecture or auditing, while other models handle parallel tasks. Google says the agents can work in separate workspaces and use additional agents to check changes.
A model aimed at longer, more complex tasks
Google says Carbon uses more tokens than Argon to explore problems, reason through them and run extended verification loops. That distinction matters because more extensive model work can mean higher usage costs; the listed API rates describe token prices, not the total cost of completing a task.
The company presents Argon as the better fit for most software engineering and knowledge work, reserving Carbon for tasks that push the model further. In its announcement, Google reports that Carbon ranked first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than twice lower cost per task than the next-best external model. These are company-reported benchmark results, rather than a general measure of performance across all tasks.
Google also describes a new Multimodal Knowledge Agent benchmark, developed with Artificial Analysis, for evaluating agents working with dense files that can include charts, images, technical drawings and tables. Google says Carbon ranked second on the benchmark, with average cost per task more than six times lower than that of the top-performing model.
Google’s reported safety and research results
The announcement includes results from Google’s own safety testing as well as external evaluations. In simulated cyberattack environments, Google says its safeguards blocked 99% of attack trajectories from an adaptive red-teaming agent, with a per-attempt breach rate of 0.8%, compared with 3.1% for Argon. It also reports that Carbon passed 92% of tasks on its 375-task agentic honesty benchmark, versus 70% for Argon.
Google says researchers used Carbon alongside other models in a formal-verification project involving Signal’s cryptographic protocol. The announcement describes the work as a demonstration of how model-generated Lean code could help researchers prove security properties. It also reports that, in a test involving seven low-homolog protein families from the CASP17 blind test set, Carbon increased the number of retrieved homologs by an average of 40-fold. Google says broader benchmarking of that work is still underway.
These results describe specific tests and projects, not a guarantee that Carbon will deliver the same performance in other settings. Access is also tiered: general consumer access is for Google AI Ultra subscribers, while research and cybersecurity access involves application and selection.