Google has made Gemini 4 Carbon, its most powerful artificial intelligence model, generally available to paying developers, businesses and top-tier subscribers, two weeks after first handing it to a small group of cybersecurity defenders. The wider release, announced on Tuesday, puts the company's flagship model into the hands of anyone who pays for the Gemini application programming interface or for Google's most expensive consumer subscription.
Carbon was first unveiled on 28 October, but its initial release was limited to selected defenders working through Google DeepMind's Fairwind programme, where the model was used for finding and fixing software vulnerabilities. On 10 November, Google opened it up much more widely. Developers can now call it through the paid tier of the Gemini API in Google AI Studio and Android Studio, enterprises can use it through Gemini for business, and Google AI Ultra subscribers are receiving access in the Gemini app, AI Mode in Search and the company's Antigravity coding environment.
Koray Kavukcuoglu, senior vice president of Google DeepMind and Google's chief AI architect, described Carbon in the announcement as the company's "strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows". Google says Carbon builds on Gemini 4 Argon, which reached developers on 13 October, but uses more reasoning tokens to investigate problems, evaluate possible solutions and run additional rounds of verification before answering.
The model is aimed squarely at the hardest kind of work AI systems are now asked to do: long, multi-step tasks where a request requires planning, execution, testing and refinement, rather than a single short answer. Google describes it as its strongest model for autonomous agents, complex knowledge work, software engineering and scientific reasoning, and says it exceeds Argon on tasks that test the limits of model intelligence.
Pricing is set at $2 per million input tokens and $10 per million output tokens during an introductory period that runs until 30 November. From 1 December, the rates rise to $4 per million input tokens and $20 per million output tokens. The free tier of the Gemini API does not get Carbon, and neither does a Google AI Pro subscription; consumers need Google AI Ultra, which costs $99.99 a month for the Ultra 5x plan and $199.99 a month for the Ultra 20x plan, which offers higher usage limits.
Alongside the model itself, Google is previewing a feature called /swarm, which coordinates teams of AI agents to split complex tasks between them in real time. The idea is that a single prompt can set off a group of agents that divide the work, check each other's output and bring the results together, rather than leaving one model to grind through a long task alone.
Google is also expanding access for scientific research. Eligible researchers can apply to use Carbon and Argon through Antigravity's highest usage tier without charge for six months, subject to selection, capacity and programme requirements. The company has not published a country-by-country rollout schedule, and has not announced downloadable model weights or a self-hosting licence: Carbon is a proprietary, hosted model, available only through Google's products and API.
The launch comes at a crowded moment in the AI race. Carbon's wider release arrived a day after Anthropic launched Claude Fable 5.2 and Mythos 5.2, and the same week Anthropic began the roadshow for its blockbuster stock market listing. Google's own announcement notes that Carbon was, until now, available only to defenders; the company is clearly keen to show that its strongest model can compete in the open market, not just in a gated security programme.
Independent rankings tell a mixed story about where Carbon stands. The model had topped some leading benchmarks when it was announced, but Anthropic's new models knocked it off the top of at least one widely watched independent ranking the day before Carbon's general release. Google has not published fresh benchmark comparisons with its latest announcement, and the company says its own reported scores should be treated with the usual caution until independent testing catches up.
For developers, the practical question is what Carbon can do that cheaper models cannot. Google's pitch is that the extra reasoning is worth the premium price for work where a mistake is expensive: auditing code for security flaws, running long research projects, or carrying out multi-stage engineering tasks where each step depends on the last. For everyday questions, the company itself concedes, most people will not need it — Kavukcuoglu wrote in the announcement that most users will be well served by lighter models.
What happens next will depend on how quickly developers adopt it. The introductory pricing runs only until the end of the month, so the first real test of demand will come in December, when the price doubles. If Carbon finds its niche in security work and long-horizon agent tasks, it could become the model Google's most demanding customers reach for first; if not, it risks becoming another flagship that is admired more than it is used.
