Google has made Gemini 4 Carbon available to paying developers and enterprises, and is rolling it out to Google AI Ultra subscribers. In a 10 November post, Koray Kavukcuoglu, Google DeepMind’s chief AI architect, calls Carbon the company’s strongest model, with “frontier performance in cybersecurity capabilities and long-horizon agentic workflows.”
Google CEO Sundar Pichai said: “Gemini 4 Carbon is rolling out today. It’s built for sustained, long-horizon tasks, reasoning deeply through complex software engineering, cybersecurity, and scientific discovery.”
Google first announced Carbon on 28 October, and at first opened it only to a subset of cybersecurity defenders through its Fairwind Program. Today’s post is the wider product rollout. It is a different model from Gemini Robotics 2.5, the vision-language-action system DeepMind described separately. The Verge, 9to5Google and VentureBeat have also reported the release.
Where you can use it
The company says Carbon is available through the paid Gemini API in Google AI Studio, Android Studio and Gemini for business for all paying developers and enterprises. Google AI Ultra subscribers can use it in the Gemini app, AI Mode in Search, and Google Antigravity. Enterprises can also try it in CodeMender. The posts do not say it is rolling out to free or standard Gemini Pro plans.
On Antigravity, the official account says Carbon is live in the desktop app, the CLI and IDE extensions for Ultra subscribers, and in Gemini for business. “We are working on expanding access to more paid users soon.” Scientists and researchers can still apply for six months of Antigravity’s highest usage tier, now including Carbon as well as Gemini 4 Argon, subject to eligibility, capacity and selection. Cybersecurity defenders can still apply to Fairwind for “Carbon for cyber.”
Google is also previewing a /swarm command in Antigravity, which it says turns a single prompt into a coordinated team of agents. In that setup, Carbon can be assigned to roles that need the deepest reasoning, such as architectural design, coordination and auditing, while faster, lighter models work in parallel isolated workspaces, with “adversarial verifiers that test every change.” In one official demonstration, a team of Carbon agents used /swarm to turn historical photos and engineering plates into an interactive Three.js exhibition of the Eiffel Tower’s construction, from a single prompt. This is a separate change from Antigravity’s recent Knowledge Work layout, which hides Git and the terminal for research and writing.
How Google positions Carbon against Argon
Gemini 4 Argon, launched in mid-October, “remains the ideal choice for the majority of software engineering and knowledge work,” Kavukcuoglu writes. Carbon is “built for tasks that test the limits of a model’s intelligence,” such as complex multi-step coding and debugging, and open-ended scientific research. The company says it uses more tokens than Argon “to explore thoroughly, reason deeply, and run extended verification loops.”
Price, and what Google claims it can do
API access is priced at the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. After 30 November, the company says that rises to $4 and $20. Google Cloud’s pricing page lists the same global figures, with slightly higher non-global rates.
Today’s post says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, “at more than 2X lower cost per task compared to the next best external model.” Google also says it ranks second on Artificial Analysis’s new Multimodal Knowledge Agent benchmark, at “more than 6X lower in average cost per task compared to the top performing model.” When Carbon was announced in October, Google said it ranked first on Artificial Analysis Intelligence Index v5.0, with a score of 59, and reported 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, evaluated without guardrails. Those are Google’s cited scores, not independently re-run here.
The company describes Carbon as stronger at knowledge-work artefacts such as spreadsheets, presentations and reports, at full-stack and 3D web work, and at scientific tasks. It also says a “Carbon for cyber” path, reached through Fairwind, is meant for finding, validating and patching vulnerabilities. Early examples it cites, including work with Wiz’s Scan for Good programme and Google’s own Cloud Vulnerability Research team, are Google’s and its partners’ accounts; some vulnerability details, it says, will be shared after responsible disclosure.
On safety, Google says it has strengthened misuse safeguards as the model’s cyber capabilities have grown. It reports, among other internal and external tests, a 0% attack success rate on Gray Swan’s indirect prompt-injection benchmark, while also saying prompt-injection robustness “remains an area for continued improvement.”
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
