Google on Tuesday, Nov. 10, opened Gemini 4 Carbon to developers, businesses and some paying consumers. Google calls Carbon its strongest model and a frontier model, the term for the most capable AI models, ones that match or beat the best that exist today. In late October, Google had first released it to a set of cybersecurity defenders through its Fairwind Program.
Developers can build with it through the paid tier of the Gemini API, a connection that lets developers plug a service into their own apps. It is not on the API's free tier. Businesses get it through Gemini for business. It is also rolling out to subscribers of Google AI Ultra, Google's top consumer plan, which costs from $99.99 a month in the U.S., in select countries. They can use it in the Gemini app, in the chat-style AI Mode in Google Search, and in Google's tools for building software, AI Studio and Antigravity. The $4.99-a-month Plus and $19.99 Pro plans don't list it, and Google has not said whether they will get it. Scientists can apply for six months of free access through Antigravity, if Google selects them.
Prices are set per million tokens, the small chunks of text, often parts of words, that AI models read and write. Through Nov. 30, developers pay $2 per million tokens Carbon reads and $10 per million it writes, the same as for Gemini 4 Argon, Google's faster model that spends fewer tokens per answer. From Dec. 1, the price for both models rises to $4 and $20. Google says Argon is still the better choice for most work, and pitches Carbon at the hardest tasks, where it uses more tokens than Argon to work through a problem and check its answers.
Google says Carbon ranks first on tests of coding, research and business tasks that Cursor, Perplexity and Zapier built, at less than half the cost per task of the next best model from outside Google. Vals AI, an independent company that tests AI models on finance, legal, tax and coding work, ranks it third of 36 models at 70.45 percent. Anthropic's Claude Fable 5.2 leads at 71.31 percent and Claude Sonnet 5.5 is second at 70.85 percent. Vals shows each score with a range of about one point either way, and the top three ranges overlap, so the order among them is not settled. Apart from the Vals ranking, the results so far come from Google or from companies that tested Carbon early.
Google is also rolling out an early version of a feature called /swarm in Antigravity, its app for building software with AI agents, AI that can carry out tasks on its own. One prompt starts a team of agents: a coordinator, workers that write code side by side, and checkers that run tests and challenge each change. Google suggests Carbon for the coordinator and checker roles, with lighter, faster models doing the routine edits. It warns that such teams use up usage limits much faster than a normal chat.
Google says Carbon is more capable at computer security than its earlier models, so it strengthened safeguards meant to stop attackers from misusing it. In Google's own simulated attacks, by an automated attacker that changes its plan after each block, Carbon stopped 99 percent of attack runs. The share of single attempts that got through fell to 0.8 percent, from 3.1 percent for Argon. Security teams can still apply to Google's Fairwind Program for a version called Carbon for cyber, which Google says leads key tests of finding, exploiting and fixing software flaws.