Google has opened its strongest model yet, Gemini 4 Carbon, to paying developers and enterprises, and it is pairing the release with a preview of a feature that could change how teams hand off software work. Antigravity's new /swarm command turns a single prompt into a coordinated group of AI agents that can be tracked in real time. The launch arrived on November 10, 2026, with an introductory price that Google says will rise after November 30.
What Google Launched
Google positions Carbon as its most capable model, built for harder work than everyday coding and documentation. The company says its earlier model, Argon, remains the ideal choice for most software engineering and knowledge work, while Carbon targets tasks that call for deeper reasoning and more verification. That split matters for teams deciding where to route requests. Faster models can handle volume, and Carbon is best reserved for the steps where a wrong answer is expensive.
Access is tiered. Developers and enterprises can call the model through the paid Gemini API in Google AI Studio and through Gemini for business. Google AI Ultra subscribers can use it in Google Antigravity, the Gemini app and AI Mode in Search. A separate Fairwind program gives selected cybersecurity defenders a version with fewer safeguards. Google has not announced downloadable weights or a self-hosting option, so Carbon remains a hosted service.
Pricing and the December Change
The headline rate is $2 per million uncached input tokens and $10 per million output tokens. That matches the introductory price for Argon and runs through November 30. After the promotion, the rate for the public Carbon model rises to $4 per million input tokens and $20 per million output tokens.
Cached input receives a 95 percent discount, which works out to $0.10 per million tokens during the promotion and $0.20 afterward, according to coverage of Google's pricing. Teams should model both periods. A workload that looks affordable in November could cost roughly twice as much per token in December if its volume and mix stay the same.
Inside the /swarm Preview
The more unusual part of the launch is /swarm, a preview in Antigravity that turns one request into a team of agents. Carbon handles high-reasoning roles such as architecture, coordination and auditing. Lighter, faster models run in parallel inside isolated workspaces, and adversarial verifier agents test each change before it is accepted. Users can follow the team's progress as it works.
Google demonstrated the feature with projects including a three-dimensional Eiffel Tower exhibition, a 3D reconstruction of an office and an open-world browser game. Those are showcases rather than independent tests. Google has not published separate pricing for swarm runs, and a session with many parallel workers will consume more tokens than a single chat, so budgets should be set accordingly.
The idea to watch is per-role model assignment. A team can put expensive reasoning where judgment matters and cheaper models where the work is repetitive, rather than paying frontier rates for every step of a job.
Benchmarks and Safety Claims
Google reports that Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, all coding and agent benchmarks, at more than twice lower cost per task than the next-best external model. On a multimodal knowledge-work evaluation run with Artificial Analysis, Google says Carbon ranks second at more than six times lower average cost per task than the top model. These are company-reported results, and independent leaderboards will take time to confirm them.
Safety claims are also central. Google says Carbon blocked 99 percent of attack trajectories from an adaptive red-teaming agent in simulated cyberattack tests, and it reports a 0 percent prompt injection success rate on the Gray Swan indirect prompt injection benchmark. For agent builders, injection resistance is the figure that matters most, and it should be retested against their own tools and data before anything goes to production.
Why It Matters
Agent products are moving from single assistants toward teams of cooperating agents, and pricing will decide which teams can afford to run them. A swarm that calls a frontier model for every subtask is expensive. One that reserves frontier reasoning for planning and review is a different economic proposition. Google's launch makes that routing choice explicit inside the product.
The December price increase also sets a clock. Developers who want to test Carbon at introductory rates have a few weeks. The safest approach is to measure cost per completed task now, while the rate is lower, and then decide whether the new price still fits the work.
