How Much Does Gemini 4 Carbon Cost and Who Can Use It?

How much does Gemini 4 Carbon cost?

Gemini 4 Carbon costs $2 per million input tokens and $10 per million output tokens through the Gemini API until November 30. That is an introductory price. The standard price is $4 per million input tokens and $20 per million output tokens, so the cost doubles after the intro period ends. Cached input is discounted by 95%. You can call Carbon through the Gemini API today. Google AI Ultra subscribers also get it inside Google's Antigravity platform.

A token is a small chunk of text, roughly three-quarters of a word. Prices are quoted per million of them, and you pay separately for what you send in and what the model writes back.

What happens to the price after November 30?

The introductory rate ends on November 30, and the price then doubles to $4 per million input tokens and $20 per million output tokens. That matches the standard price of Claude Opus 5.5, according to one reviewer who covered the launch. He also noted that Argon launched at the same $2 and $10 intro rate.

Nobody outside Google knows what happens next. The same reviewer guessed that Google might extend the intro price or release a new model at a similar price point. That is a guess, so don't build a budget on it. Plan for the $4 and $20 rate.

Here is how Carbon compares with the two Anthropic models mentioned in the coverage:

At the intro rate, Carbon costs a quarter of what Fable 5.2 charges per token. After November 30 it costs half.

Who can use Gemini 4 Carbon right now?

There are two doors, and they have different rules.

Through the Gemini API. Developers can call Carbon today and pay per token. This is the route for anyone building Carbon into a tool, script or internal app.

Through Antigravity. Antigravity is Google's development environment, where you hand the model a task and it works through it. Google AI Ultra subscribers get Carbon there. One reviewer described Ultra as the highest and most expensive tier Google offers. The sources don't give the Ultra subscription price, so check Google's own page for the current figure.

If you are not a developer and don't pay for Ultra, the sources don't describe a route for you to use Carbon yet. A Flash-sized Gemma model has reportedly been signaled as coming, which may be the more affordable option for everyday use. That is not confirmed in detail, so watch Google's announcements.

What are the thinking levels and context limits?

Carbon has three thinking levels: low, medium and high. Medium is the default. A higher level means the model reasons longer before it answers. That usually helps on hard problems and costs more time and tokens.

Google has also removed some controls that developers used to have. With the Gemini API you no longer have control over things like temperature, top P and top K, the settings that tune how random a model's answers are. Google presets them. Reasoning effort is the main dial you get.

The limits:

• Context window: about one million tokens. That is roughly 750,000 words of material the model can hold in view at once, by the usual rule of thumb.
• Output limit: up to one million tokens, the same as Argon.
• Inputs: one reviewer described the supported input types as "basically everything," meaning text, images and more.

A big context window does not mean you should fill it. You pay for every input token on every call, so a long document sent repeatedly adds up. That is where the 95% cached input discount helps. If you send the same large block of material many times, cached tokens cost a small fraction of the normal rate.

How much will a real task cost?

Per-token prices are hard to picture, so here is the arithmetic at both price levels.

A single request with 100,000 tokens in and 10,000 tokens out:

• At the intro price: $0.20 for input plus $0.10 for output, so $0.30.
• At the standard price: $0.40 plus $0.20, so $0.60.

A long agent run with 5 million tokens in and 500,000 tokens out:

• At the intro price: $10 plus $5, so $15.
• At the standard price: $20 plus $10, so $30.

Models that reason before answering typically bill their thinking tokens as output tokens, and output is the expensive side. Carbon thinks heavily. Artificial Analysis measured about 26 minutes per task on its index, against a median of about 13 minutes, and about 205 million tokens against a median of about 72 million. Expect it to be thorough and slow. One reviewer said he launches a Carbon task and comes back later rather than using it for quick chat.

For a sense of scale, one reviewer said Gemini reported using 26 million tokens across multiple agents to build a browser-based Excel clone, in about an hour and 40 minutes end to end. He called it "a couple dollars." That figure came from the model's own report, not a bill, so treat it as a rough signal.

Another reviewer left Carbon running on a 3D reconstruction task, and after 21 hours it was still researching, building and reviewing its own work. Long runs like that are what the token meter punishes.

Is Carbon worth the price compared with Fable 5.2?

It depends on the task. Fable 5.2 leads the Artificial Analysis index at 65, and Carbon sits second at 60. A reviewer who tracked the index put Fable 5.2 at about $26.50 per task and Carbon at a little under $7.80. For five fewer points, you pay roughly a third per task.

Reviewers disagreed on how much the gap matters. One said that for everyday work the cost per task is the number that counts. Another said that Carbon is a good, reliable model but "nowhere near" Fable 5.2 in his own daily use, and he kept going back to Fable. Both can be true. Benchmarks measure averages. Your work might sit at either end.

A sensible approach is to match the model to the job:

• Use Carbon for long background tasks where the cost per task matters. Code security reviews, large document analysis and overnight builds are examples.
• Use a top-priced model when a mistake is expensive and you want the strongest judgment you can buy.
• Test both on three or four real tasks from your own work before committing.

What should you do before November 30?

1. Decide whether you need Carbon at all. If you only chat with AI and don't write code or call APIs, you may not need it.
2. Run a small trial at the intro rate. Pick tasks you already do, run them at medium thinking, and record the tokens used and what you were charged.
3. Budget at the standard rate. Double your trial cost to see what the same work costs from December.
4. Compare on cost per finished task. A cheaper model that needs three attempts can cost more than an expensive one that needs one.
5. Use caching. If your prompts share a large common block, cached input is 95% cheaper.

Could you build your own AI spend tracker?

Once you are comparing per-token prices across models, a spreadsheet gets clumsy. A small internal tool could log each task, the model used, the tokens consumed and the result, then show cost per finished task for your team. You can describe an app like that in plain English to Remy, which builds and deploys full-stack apps, including the database and sign-in. It bills AI tokens at cost with no markup, so the tool's own running cost stays visible. Remy is a paid platform ($99 per month, or $79 billed annually), so it makes the most sense if you plan to build more than this one tracker.