Google said on Tuesday 10 November that it had opened Gemini 4 Carbon to all paying developers and enterprises. Two weeks earlier, Google had given the model only to selected cyber defenders through its Fairwind programme.

Carbon does not replace Gemini 4 Argon. Google said Argon remains the best choice for most coding and knowledge work, while Carbon is meant for the hardest tasks. Carbon uses more tokens than Argon, so The Decoder expects the bill to be higher in practice despite the same per-token price.

Fable 5.2 overtook Carbon on a benchmark index

Carbon was the top frontier model for about ten days until Anthropic released Claude Fable 5.2 on 9 November, Latent Space reported. The table shows Artificial Analysis index scores and cost per task as reported by Latent Space and The Decoder.

Artificial Analysis Intelligence Index v5. Costs as reported by The Decoder; Carbon’s figure is calculated at the December prices of $4 / $20.

The index says nothing about individual tasks: in Latent Space’s roundup, Carbon was ahead of Fable 5.2 on the Terminal-Bench 4.0 benchmark, for example (65 per cent against 60 per cent).

Google said Carbon ranks first on the CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench benchmarks at less than half the cost per task of the next best external model. Latent Space, however, reported that Perplexity announced a WANDR record for Fable 5.2, and that Cursor named Fable 5.2 (max) the best model on CursorBench with 60 per cent. Latent Space does not say which version of CursorBench the figure refers to; in Google’s chart published by The Decoder, Carbon is ahead of Fable 5.2 on CursorBench 4.0.

Agent swarms in Antigravity

Google also introduced a preview of the /swarm feature in its Antigravity development environment. The feature turns a single prompt into a team of agents whose work can be tracked in real time. Developers can give Carbon the roles that need deep reasoning, such as architecture design, coordination and review. Lighter models work in parallel in separate workspaces or test every change. The Decoder said the cost benefit of agent swarms is disputed.

Google said Carbon leads in tests that measure finding, exploiting and fixing vulnerabilities. Google itself limits this result to the version available through the Fairwind programme, and the company gives no figures.

Gray Swan’s test tries to steer the model through instructions hidden in material it reads. Google said no attack succeeded, but admitted that its defences still need improvement. The Decoder said the test may already be too easy.

Availability and price

Paying developers can use Carbon through the Gemini API in Google AI Studio, Android Studio and Gemini for business. For Google AI Ultra subscribers, the model is coming to Antigravity, the Gemini app and AI Mode in Search. Until 30 November the introductory price is $2 per 1 million input tokens and $10 per 1 million output tokens, then $4 and $20. The prices apply to the model intended for public use. Researchers can apply for six months of free access to Carbon and Argon in Antigravity, and cyber defenders can apply to the Fairwind programme for the Carbon for cyber version. Researchers’ access depends on eligibility, capacity and selection.

In GitHub Copilot, Carbon is billed by usage at the model provider’s list price, and it is rolling out gradually to Pro+, Max, Business and Enterprise users. In Business and Enterprise organisations the model is enabled by default unless an administrator has turned off default access to new models or blocked the model in Copilot’s model policies.