Gemini 4 Carbon is open to paying developers. The intro price ends Nov 30.

How that compares. Anthropic's price list (opens in a new tab) has Claude Opus 5.5 at $4/$20 and Claude Fable 5.2 at $8/$40. So "Carbon undercuts Opus 5.5" holds only through Nov. 30. From Dec. 1 the two have the same list price per token. "A quarter the price of Fable 5.2" becomes half. Per-token price is not cost per task either: models use different numbers of tokens for the same job.

What the outside numbers say. On the new Multimodal Knowledge Agent benchmark, which Artificial Analysis launched with Google DeepMind (opens in a new tab) today, Carbon (high) scored 54%. Claude Fable 5.2 (max) led with 57%, and Claude Opus 5.5 (max) scored 53%. Average cost per task: $6.29 for Carbon against $31.97 for Fable 5.2. Artificial Analysis calls that "over 5x". Google's blog says "more than 6X". Cursor says (opens in a new tab) Carbon at high effort is the top model on its own CursorBench at 60.4%. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best outside model. That claim is Google's.

What else shipped. Google is previewing /swarm in the Antigravity app and CLI. It splits one prompt across a team of agents, and each agent can be pinned to a different model. Google still recommends Gemini 4 Argon for most coding and knowledge work, and pitches Carbon for long, multi-step jobs where it "uses more tokens than Argon."

Who it's for. Teams running long agent jobs, such as big migrations, debugging across many files, research or security review, where a few extra points matter more than speed.

For builders on Google Cloud. Google's Cloud pricing page says that from Nov. 13, if a Gemini 4 completion runs past 32K tokens, you pay for the tokens produced even if the request fails on a client or server timeout. Set timeouts with that in mind.

The caveats. MODEL has not tested Carbon. Most benchmark numbers here come from Google, from Cursor's own benchmark, or from a benchmark Google DeepMind built with Artificial Analysis. A country-by-country rollout list has not been published. Consumer access requires Google AI Ultra.

Verdict. Worth a run on your hardest agent task before Nov. 30, while it is half price. Budget at $4/$20 when you decide whether to keep it, and check what it costs per finished task, not just per token.

Not tested by MODEL. Not sponsored. No affiliate links.

MODEL Editors