Gemini 4 Carbon opens to all: Sonnet prices now, Opus prices December 1

On November 10, 2026, Simon Willison released version 0.36 of llm-gemini, the plugin that hooks Google’s Gemini family into his LLM tool. The reason for the update is simple: Gemini 4 Carbon is now open to everyone. Google had announced it two weeks earlier for a small circle of cyber-defence specialists.

Defenders first, everyone else second

According to Willison, Carbon is the best Gemini Google has ever shipped. That’s the verdict of a seasoned tester, given in a short post, and for now it’s the only source we have. The post has no benchmark figures, either against OpenAI’s models or against Anthropic’s. So take it as a first-hand impression, not a spot on a leaderboard.

On the practical side, the model offers three levels of thinking effort: low, medium and high. That’s the dial that decides how long it chews on a problem before answering. It defaults to medium.

Sonnet prices, then Opus prices

Pricing is where the comparison with the competition is sharpest. At launch, Carbon costs $2 per million input tokens and $10 per million output tokens: exactly the price of GPT-6.1 Sol and Claude Sonnet 5.5. But that price has a use-by date. On December 1 it is set to double, to $4 and $20, which puts it level with Claude Opus 5.5.

The message is hard to miss: Google is shelving its model in the premium aisle and giving anyone who wants to check three weeks of launch pricing. It’s the introductory rate on a broadband contract, API edition: enjoy it now, because the letter announcing the real price is already in the post. If your team is thinking about switching, the maths takes about four seconds: the same workload will cost twice as much in December as in November. Testing will never be this cheap again.

Pelicans and a giant robot in Yonkers

Willison’s usual test is to ask the model to draw a pelican riding a bicycle. He got three, one per thinking level. But he admits it himself: Gemini’s pelicans have been excellent for a while, and the test doesn’t tell models apart much any more. So he reused a prompt he had tried the day before with Claude Fable 5.2, in Antigravity, Google’s IDE. It’s a three.js animation in which Yonkers rises up as a giant robot and stomps across Manhattan, Queens and Brooklyn, built from OpenStreetMap map data.

The result showed up, with one caveat: it took a second request to get a self-contained HTML file with no external dependencies. Willison recorded a video of it. What his post doesn’t say is which of the two models did the better job.

/swarm, or the art of delegating

Willison also points to a demo he spotted on YouTube. Using Antigravity’s /swarm feature, Carbon reportedly built an open-world 3D game. The player explores Google’s Mountain View campus as a pelican riding a GBike. To pull this off, the model coordinates sub-agents. They model the 3D objects in Blender from references produced by Omni and Nano Banana. They write the game in Three.js with WebGL shaders. Finally, they add sound effects generated with Web Audio and music produced by Lyria.

It’s a demo, not an independent test. It shows what sub-agent orchestration can do when everything goes to plan, not what it will do to your codebase at half past five on a Friday.

If you want to know whether Carbon is really worth Opus money, you have until December 1 to try it at Sonnet prices. After that, your invoice runs the benchmark, and it doesn’t grade on a curve.