Gemini 4 Carbon is finally out

Expand

Open original image ↗

Gemini 4 Carbon is finally out. We picked up the release this afternoon, and Google now lists gemini-4-carbon as generally available on the paid Gemini API. After this morning’s billing entries, this is the bit I’ve been waiting for: an actual released model we can put through our prompts. 1 2

The app rollout is starting with Ultra subscribers, as we shared in our launch-day check. That restriction is separate from the developer route: the API has its own paid tier, and Google links straight to Carbon in AI Studio. It isn’t available on the API’s Free tier. 3 4

The introductory price is $2 in, $10 out

Google’s current price table confirms $2 per million input tokens and $10 per million output tokens through 30 November. Thinking tokens are included in the output charge. From 1 December, those rates rise to $4 and $20. Cached input is $0.10 during the introductory period, then $0.20. 4

That clears up the prices in our earlier billing report. The $4/$20 catalogue entries were the standard rates, while the launch offer is lower. Carbon supports Low, Medium and High thinking, with a unified 1,048,576-token context limit. Normal synchronous output is capped at 65,536 tokens; the separate long-decode route can go up to one million. 5

Carbon High starts at 60.4% on CursorBench

There are already three new results on the official CursorBench 4.0 board. Carbon High is at 60.4%, Medium at 60.1% and Low at 58.6%. Fable 5.2 Max is at 60.0%, with Opus 5.5 Max at 57.8%. The tiny gap between Carbon High and Fable Max isn’t enough to call a decisive winner, but it’s a strong first result. 6

Expand

Open original image ↗

The cost comparison is interesting too. Carbon High is listed at $9.22 per task, against $32.79 for Fable Max. Medium is a little cheaper again at $8.49. Low costs more than either in this particular evaluation, so lower effort doesn’t automatically give you the cheapest completed task. These are Cursor’s API-equivalent task costs, rather than subscription prices.

Expand

Open original image ↗

The source chart above is the firmer comparison for now. We don’t yet have a matching Carbon result in the pinned Terminal-Bench 4.0 cohort, so no combined Coding result for Carbon is fully observed.

A couple of our earlier Carbon demos

We’ve had some good early Carbon outputs in Arena. The helicopter is a nice place to start: it builds the scene around an engineering-style interface, with the aircraft easy to inspect. These clips predate today’s API release, so I’m keeping them separate from the scored results we’ll get from the public model. 7

1:37

The seaplane is the other one I’d open. It has an ocean, islands and driving controls, and gives a much better impression in motion than a still does.

0:50

Now we can try the released version on the same creation prompts as Fable 5.2. We haven’t scored public-API Voxel, 3D or GameDev runs for Carbon yet. Those results will follow once the runs are finished.