Gemini 4 Carbon Opens to Everyone at $2/$10 per Million Tokens, but the Price Doubles on December 1

Some models are judged by benchmarks, others by a task like "make something weird and let's see what happens." Simon Willison, the author of the LLM command-line tool, traditionally takes the second route. Alongside the release of the llm-gemini 0.36 plugin, he published his first impressions of Gemini 4 Carbon, Google's new flagship model, which has just become available to everyone.

What happened and what it costs

Two weeks ago, Google announced Carbon only for a "subset of cyber defenders," that is, a narrow circle of security specialists. Now access is open to everyone else. Willison calls it the best Gemini ever.

Launch pricing is $2 per million input tokens and $10 per million output tokens. That is on par with GPT-6.1 Sol and Claude Sonnet 5.5. But there is a catch: from December 1 the rate doubles to $4 and $20, the same as Opus 5.5. In effect, Google first sells Carbon as a mid-priced model and then moves it into the premium tier.

The model has three levels of "thinking effort": low, medium and high. Medium is enabled by default. For each level, Willison generated an image for his signature test, an SVG of a pelican riding a bicycle. The high variant came out neat: a monocle, a sailor cap with a blue bow, a basket of fish, smooth gradients. The author himself admits that Gemini's pelicans have been good for a while, so this test no longer distinguishes much.

A tougher test: a city-sized mech

So Willison picked a harder task. The day before, he had given the same prompt to Claude Fable 5.2: build a three.js animation in which the city of Yonkers rises up as a giant mech robot and stomps across Manhattan, Queens and Brooklyn, using OpenStreetMap data. He ran everything in Google's Antigravity IDE.

It worked, but not right away. To get an HTML file with no external dependencies, he had to separately ask it to "bundle everything into a standalone HTML." A small thing, but telling: the agentic environment produced a project that can't simply be opened and sent to a colleague. Willison recorded his screen and posted the video.

In an update, he added a clip from Google itself. According to the description, using the /swarm feature in Antigravity, Carbon built a 3D game in which the player explores Google's Mountain View campus as a pelican riding a GBike bicycle. The model coordinated subagents. Some modeled 3D assets in Blender from references produced by Omni and Nano Banana, while others wrote a browser game in Three.js with WebGL shaders. Procedural sound effects via Web Audio and music from Lyria were added as well. This is a vendor demo, not an independent test, but it shows the target scenario: orchestrating several subagents and different multimodal models within a single task.

What to do with this

For teams watching their budgets, the takeaways are as follows:

• Test now. If Carbon fits your task, tokens are half the price until December 1. After that date, you should compare it with Opus 5.5, not with Sonnet-class models.
• Thinking mode is a lever. Medium is the default. Check where low is enough and where high really delivers a noticeable gain. Both quality and the final bill depend on it.
• Plan for a "packaging" step. An agentic result may need an additional prompt to become a self-contained artifact, as in the HTML file story.
• For a quick try from the terminal, update the llm-gemini plugin to 0.36. It was released at the same time as the model's opening.

Limitations

Everything we know is the impressions of a single developer on a few prompts. The post has no benchmark figures, response speed or stability data. There are also no details on cybersecurity, the reason access was restricted at first. So run Carbon on your real tasks while the introductory price lasts, and only then decide whether to move production to it.