Google has made Gemini 4 Carbon available to developers, enterprises, researchers and consumers, positioning the model above Gemini 4 Argon for tasks that require longer reasoning and repeated verification. The company says Carbon is designed for complex software work, scientific investigation, cybersecurity and document-heavy professional workflows rather than the broader everyday workloads assigned to Argon.
Carbon is available through the Gemini API in Google AI Studio and through Gemini for business for paying developers and companies. Google set an introductory price of $2 per million input tokens and $10 per million output tokens, matching the introductory rate for Argon. The model is also rolling out to Google AI Ultra subscribers through the Gemini app, AI Mode in Search and the company’s Antigravity development environment.
The announcement emphasizes “long-horizon” agentic work: projects in which a model must plan, use tools, inspect results and repeat checks across many steps. Google says Carbon spends more tokens than Argon to explore and verify difficult problems. That design may make it better suited to architectural decisions, multi-package code migrations and open-ended research, but customers will need to weigh the additional computation and latency against the value of deeper analysis.
Google is also previewing a /swarm feature in Antigravity. It can turn one instruction into a group of agents working in separate environments, with Carbon assigned to coordination, design or auditing while faster models handle narrower jobs. The company demonstrated the approach on 3D reconstruction, browser development and creative-coding projects. Those demonstrations are company-selected examples rather than independent guarantees of performance.
For cybersecurity, Google says Carbon can inspect software, identify and chain flaws, and propose root-cause patches. Access for defenders began two weeks before the broader release through the Fairwind Program. Google reports leading results on several security evaluations covering vulnerability discovery, exploitation and repair, though real deployments will require their own controls and validation before generated findings or patches can be trusted.
The company also cited first-place results on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, along with lower claimed cost per completed task than competing external models. In multimodal document work, Carbon ranked second on a new benchmark developed with Artificial Analysis, according to Google. Benchmark rankings remain dependent on test design, tool configuration and pricing assumptions, so they do not establish superiority for every workload.
Carbon extends Google’s strategy of offering multiple frontier models tuned for different balances of speed, cost and depth. Its practical value will depend less on isolated scores than on whether it can sustain accurate work across long projects, recover from failures and produce outputs that survive human and automated review.
