Google DeepMind made Gemini 4 Carbon available to developers, enterprises, scientists and consumers on 10 November 2026, widening access after its limited release to selected cybersecurity defenders through the Fairwind Program. The release brings the model to Google’s paid API and begins a rollout to Google AI Ultra subscribers. Google is also previewing /swarm in Google Antigravity, which lets a single request set several AI agents to work on a project.
Key points
- Carbon is available through the paid API at an introductory $2 per million input tokens and $10 per million output tokens.
- Google is previewing /swarm, which assigns agents parallel work that can be tracked in real time.
- Google claims leading results on three work-oriented benchmarks, while VentureBeat reports that Claude Fable 5.2 now leads a separate general intelligence index.
- Carbon is rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.
Antigravity divides a prompt among agents
Google describes /swarm as a way to turn one prompt into a coordinated group of agents whose progress can be followed in real time. A developer can give Carbon the work requiring deeper reasoning, including design, coordination or auditing, while lighter models take other parts of the project. The agents work in separate spaces, with additional agents checking changes. It is closer to handing copies of the same brief to people at separate desks than to making one assistant finish every stage in sequence.
Preparing a report from a stack of documents could involve work proceeding in parallel and a separate check of the result, if the coordination Google describes carries over to that kind of task. It would still depend on the agents finding the right material and checking the work they produce.
Google’s demonstration used seven public planning PDFs containing 206 extracted engineering drawing views. From one prompt, a group of Carbon agents using /swarm produced an interactive 3D reconstruction across 10 overnight loops, according to the company. The example puts the feature’s division of labour on a concrete project, while the release describes /swarm itself as a preview.
Carbon’s wider launch follows its 28 October unveiling, when access went first to selected cybersecurity defenders. At that point, AI Affairs reported Google’s claims that the model led an AI index and cybersecurity benchmarks. The 10 November release opens access to the general-use model beyond that initial group.
Carbon’s API starts at Argon’s introductory price
Paying developers and enterprises can access Carbon through the API in Google AI Studio and Gemini for business. Google sets the introductory rate at $2 per million input tokens and $10 per million output tokens, matching Gemini 4 Argon. Input tokens cover material sent to the model; output tokens cover what it generates. A project that repeatedly asks agents to inspect, revise and check work therefore uses tokens on more than the final answer.
VentureBeat reports post-introductory rates of $4 per million input tokens and $20 per million output tokens. It also points out that a lower price per token can still produce a higher bill for a completed task when a model needs more tokens, retries or human correction. That distinction matters particularly for the extended verification loops Google says Carbon is designed to run.
Carbon is a proprietary hosted model distributed through Google’s products and API, with no downloadable weights or self-hosting licence announced, VentureBeat reports. For companies building on it, the available route is consequently through Google’s hosted access. Google says Carbon is rolling out to Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search, alongside the paid API access for developers and enterprises.
Carbon’s benchmark results vary by task
Google says Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than 2X lower cost per task than the next best external model on those evaluations. Those are Google’s reported results for the named tests of work carried out by agents. On Artificial Analysis’s general intelligence index, Claude Fable 5.2 has moved ahead of Carbon, which debuted there in first place.
The task-specific results also differ. HackerNoon reports that Carbon scored 59 on the Artificial Analysis Coding Agent Index, behind Claude Code running Opus 5.5 at 62. Google positions Carbon for complex coding and debugging that can involve several stages of work, while saying Argon remains its choice for most software engineering and knowledge work.
Google also worked with Artificial Analysis on a Multimodal Knowledge Agent benchmark, which tests agent tasks involving dense files containing material such as charts, images and tables. Google says Carbon ranked #2 on that evaluation and had an average cost per task more than 6X lower than the top-performing model.
