How to use Gemini 4 Carbon

Where you can use Gemini 4 Carbon

Carbon is available in Google’s own products, in two big coding tools and on the API. Google limits the consumer apps to its top AI Ultra plan; developers need the paid tier of the Gemini API, because Carbon is not on the free tier.

Why Gemini 4 Carbon may not work for you

No free tier. Google does not offer Carbon on the Gemini API free tier, so a free AI Studio key cannot use it. Turn on billing or use a paid gateway.

Not in your Gemini app. In the consumer apps Carbon is only for Google AI Ultra subscribers; other plans do not see it.

Region. Google’s list of regions for the Gemini API and AI Studio does not include mainland China, Hong Kong or Macao.

Gemini CLI. Gemini CLI is not tuned for the Gemini 4 series yet. Use Antigravity, Copilot, Cursor, or any agent that can call an OpenAI-compatible or Gemini endpoint.

Timeouts and cut-off answers. Carbon can think for minutes before it writes anything; Artificial Analysis measured about 68 seconds to the first answer token on average. Short client timeouts end the request first, and a low output limit can be used up by thinking, because the limit covers thinking plus the answer. Google recommends at least 32,768 output tokens and HTTP timeouts of 15 minutes or more, and its model card warns of occasional slowness and timeouts.

Subscription quota. Agent runs in Antigravity draw on the AI Ultra quota. One developer running four parallel sessions used up a $99.99 Ultra window in about two hours.

How to call it, and the settings that matter

On apimodels.app, call gemini-4-carbon with the same key you use for every other model: POST https://api.apimodels.app/v1/chat/completions in OpenAI format, or POST https://api.apimodels.app/v1beta/models/gemini-4-carbon:generateContent in native Gemini format. No Google Cloud project or overseas card is needed, and it works from mainland China without a proxy; top up with Alipay, PayPal, a card or USDT.

Thinking level. Carbon has three levels, low, medium and high, with medium as the default. Pass reasoning_effort on the OpenAI-compatible endpoint, or generationConfig.thinkingConfig.thinkingLevel on the native one. Higher levels spend more tokens and take longer to start.

Output limit and timeouts. Set max output to at least 32,768 tokens, give your HTTP client a 15-minute timeout, and stream long answers. Outputs beyond 65,536 tokens need Google’s long-decode mode.

Parameters it ignores. temperature, top_p and top_k have no effect on Carbon; remove them. Explicit context caching is not supported either; repeated prefixes of 4,096 tokens or more are cached implicitly.

Inputs. Carbon reads text, images, video, audio and PDFs in one request, with a 1,048,576-token context shared by input and output.

What people have built with Gemini 4 Carbon

Carbon’s public showcase is long, self-directed builds: games, 3D scenes and simulations that an agent works on for an hour or more. A few examples, all by other people and linked to their posts; the full credited list is in our Gemini 4 Carbon case library, linked at the end of this page.

A cozy diorama builder in one HTML file. FireTable ran Carbon in Antigravity’s multi-agent /swarm mode from a one-line Chinese prompt: 1 hour 41 minutes, 9 agents, 22,910 lines with no merge conflicts, procedural stone walls, 195,000 blades of GPU grass and WebAudio, shipped as a single 930 KB file.

A game rebuilt from a video, with its own bug checks. One creator gave Carbon in Antigravity the gameplay video of a 3D shooter that Claude Fable 5.2 had built in an 8-hour run and asked for the same game or better. In 48 minutes and $3.82 it produced a single 638.5 KB HTML file, and a headless-Chrome screenshot loop caught and fixed three bugs on its own. Another creator, through Vertex AI, got a playable Mario-style 3D platformer of 2,599 lines from one shot, at a reported $0.45.

Fifteen projects with every prompt published. The creator 亚里 ran 15 builds in Antigravity at high thinking, from a Counter-Strike Dust2 remake to a video editor and a rabies explainer, and published each prompt. A Minecraft-style sunset scene took Carbon one 32-minute pass where Argon had needed six rounds over about 2 hours 18 minutes. His conclusion: Carbon handles long tasks well, its visual taste is average, and one-line wishes rarely work.

Fixing what a person could not. A Hacker News user had Carbon reverse-engineer a fishing game’s data files to change spawn and hook rules, something he had failed to do for four years, and published the whole session. Another developer had it review an unfinished paper draft and reported that it raised points every other model missed.

Research. A Google DeepMind researcher reported that teams of Carbon agents found new exact solutions to the differential equations of nuclear fusion devices, the post with the most likes in our collection.

Science and simulation. A developer had Carbon write a pure CUDA and C++20 ray tracer of 65,536 raindrops falling on a lotus pond, with physics and procedural audio; another rebuilt a 181-part Song Dynasty timber pavilion with an exploded view and an ink-painting shader.

How to prompt it

Spell out the job instead of making a wish. Testers who gave Carbon one vague line got average results; the builds that worked named the stack, the deliverable and what done looks like. KostiKoNeuro’s winning racer prompt, for example, names the game, the era and the feeling, asks for music and sound in code, one index.html with libraries from a CDN, and ends with “Don’t ask questions: build it, play it yourself, fix what’s broken.”

Make it test the result, not just its own unit tests. In AICodeKing’s six builds, every one of Carbon’s final summaries overstated at least one feature, and it had only run its own tests. Ask it to open the app in a browser, click through the main flow and report what it saw.

Budget time and tokens. Carbon is slow and verbose by design: about 30 minutes per fix in one real-code test, a third more output than Argon and four times Fable 5.2’s. Run it in the background or in parallel, and cap spend with max output and a stop condition.

Use high thinking only where it pays. Medium is the default and is enough for most of the cases on this page; Cursor’s benchmark found Carbon at medium within 0.3 points of high at lower cost.

What Gemini 4 Carbon costs

Carbon has the same per-token price as Gemini 4 Argon. Per job it varies widely, because the thinking tokens are billed as output. Published examples: $0.56 per try on a hard Sudoku at high thinking; $2.12 per task on Zapier’s AutomationBench and $9.22 on CursorBench at list price; about $23 for six full app builds in AICodeKing’s test; and $27.68 for three long builds in AI for Mortals’ test, where Claude Fable 5.2 cost $883.30 for the same three.

When not to use Gemini 4 Carbon

Interactive coding and everyday edits: use Gemini 4 Argon. Google recommends it for most software engineering, it answers sooner, and it costs the same per token; the full comparison is linked below.

Short, high-volume requests: neither Gemini 4 model; Gemini 3.8 Flash is $0.45 / $2.25 per 1M tokens on apimodels.app.

When the result must be right and you cannot check it: Claude Fable 5.2 scored higher in most hands-on tests and on Artificial Analysis (65 against 60), at a much higher cost per task; that comparison is linked below too.

Polished front-end design: several testers found Carbon’s UI work average next to its 3D and long-task results.