Key takeaways
- Google opened Gemini 4 Carbon on Nov. 10 to the paid Gemini API tier, Gemini for business, Antigravity and Google AI Ultra subscribers in the Gemini app.
- The API costs $2 per million input tokens and $10 per million output tokens through Nov. 30, 2026, then $4 and $20, so every job costs twice as much.
- Google says Carbon ranks first on CursorBench 4.0 and blocked 99% of attack trajectories in simulated cyberattack tests, figures that are company reported.
In this story 4 sections
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on Nov. 10, and the Gemini 4 Carbon pricing it set for the API comes with an expiration date. Developers pay an introductory $2 per million input tokens and $10 per million output tokens through Nov. 30, 2026, after which the rate rises to $4 and $20, according to Google's Gemini 4 Carbon availability post.
The wider release comes two weeks after Google first introduced the model on Oct. 28, when only security teams in its Fairwind Program could use it. We covered that limited rollout and its benchmark claims in our earlier Carbon report.
Who can use Gemini 4 Carbon now
Google, a unit of Alphabet (NASDAQ: GOOGL), listed these ways in. The post carries the byline of Koray Kavukcuoglu, chief AI architect at Google.
- Developers: the paid tier of the Gemini API in AI Studio and Android Studio.
- Businesses: Gemini for business, through the API or the model picker, along with Google Antigravity and CodeMender.
- Consumers: the Gemini app and AI Mode in Search, but only for Google AI Ultra subscribers.
- Scientists: six months of free access to Argon and Carbon at the top usage tier in Antigravity, through an application that Google said depends on eligibility and capacity.
- Security teams: the Fairwind Program keeps adding members on a rolling basis.
"Today, we're making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers," Kavukcuoglu wrote.
The post does not state a context window size or list the countries where the model is offered.
What Nov. 30 means for API bills
Both the input and output prices double once the introductory period ends, so every API job costs twice as much after that date, whatever its mix of tokens. Google's footnote says the higher rate applies to the Carbon model intended for public access.
Our own example shows the effect. A job that sends 10 million input tokens and gets back 2 million output tokens would cost $40 at the introductory price ($20 for input plus $20 for output) and $80 at the standard price ($40 plus $40).
For a rough comparison, Anthropic launched Claude Fable 5.2 on Nov. 10 at $8 per million input tokens and $40 per million output tokens. On list price alone, Carbon's standard rate is half of Fable's, and its introductory rate is a quarter. List prices do not settle the cost of a task, though. Google itself positions its Argon model as the default for most coding and office work and steers Carbon toward harder jobs that consume more tokens.
Google's Oct. 28 post said cached input would cost 95% less than the regular input price. The Nov. 10 post does not repeat that discount.
Google's performance claims
All of the figures below come from Google and have not been checked independently.
- Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, while costing less than half as much per task as the next best outside model, the company said.
- It ranks second on the Multimodal Knowledge Agent benchmark from Artificial Analysis, with an average cost per task more than six times lower than the top model, according to Google.
- In simulated cyberattack tests, Google said Carbon blocked 99% of attack trajectories. Its breach rate per attempt was 0.8%, compared with 3.1% for Argon.
- On a 375 task honesty test, Carbon passed 92%, against 70% for Argon.
A preview of /swarm in Antigravity
Alongside the release, Google is previewing a /swarm command in Antigravity, its coding tool. It turns one prompt into a coordinated team of AI agents that users can follow in real time. Developers can give Carbon the planning, coordination and review roles while lighter, faster models work in parallel, isolated workspaces checked by adversarial verifiers. Google describes the feature as a preview and did not give a date for a full release in the post.
More coverage of new models is on our AI page.
Share this story
Spotted an error? Tell us


