Google expands Gemini 4 Carbon access, pricing it at $2 and $10 per million tokens

Google DeepMind has made its Gemini 4 Carbon model available to developers, enterprises, scientists and consumers, according to the company’s announcement on its blog. The release follows a period in which Carbon was limited to a smaller group: two weeks earlier, Google DeepMind gave cybersecurity defenders early access to the model through its Fairwind Program.

Carbon is priced at $2 per million input tokens and $10 per million output tokens, the same introductory rate Google DeepMind set for Gemini 4 Argon, the company says. The model is accessible through the API in Google AI Studio and Gemini for Business for paying developers and enterprises, and is rolling out to Google AI Ultra subscribers inside Google Antigravity, the Gemini app, and AI Mode in Search.

Google DeepMind positions Argon as the better fit for most everyday coding and knowledge work, while Carbon is aimed at harder problems: multi-step debugging, open-ended scientific research, and tasks that benefit from spending more tokens on deep reasoning and verification loops, the company says. On three independent evaluations cited in the announcement, CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, Google DeepMind says Carbon ranks first while costing more than twice as little per task as the next-best outside model.

The cybersecurity version of Carbon, offered through the Fairwind Program, leads on benchmarks that measure finding vulnerabilities (SEC-bench Pro), exploiting them (ExploitGym), and patching them (CWE-bench v1), according to Google DeepMind. Inside Google’s own Project Fortify system, the company reports Carbon reached 100% recall on scanning and 95% precision on triage. Google’s Cloud Vulnerability Research team says it used Carbon to confirm that two bugs in widely used open-source software were exploitable at a severity comparable to the recent PHP2Shell exploit, prompting maintainers to reclassify the flaw as critical. Separately, security firm Wiz says its team used Carbon to uncover previously unknown vulnerabilities, known as zero-days, in two major software-as-a-service providers in under two hours.

Google DeepMind is also previewing a feature called /swarm inside Google Antigravity, which lets a single prompt spin up a coordinated team of agents, with Carbon assigned to roles such as architectural design and auditing while lighter models handle parallel tasks.

What this means

The move signals Google DeepMind is confident enough in Carbon’s security testing to widen access beyond vetted defenders, even though the same reasoning power that finds flaws for Wiz and Google’s own vulnerability researchers could, in less careful hands, be pointed at finding them for other purposes. The company’s announcement does not say what safeguards, if any, govern broader API access compared with the Fairwind Program’s vetted cohort.

On price, matching Argon’s $2/$10 per-million-token rate for a model Google DeepMind itself calls its strongest suggests the company is betting on volume over premium pricing for its top-end model, though independent confirmation of the benchmark claims and the promised cost advantage over rival models is still limited to figures in Google’s own release.

Google DeepMind says full details of the Cloud Vulnerability Research team’s findings will be released once the responsible disclosure process with the affected software’s maintainers is complete, leaving the exact scope of the vulnerabilities not yet public.