Seongwoo Choi / Journal

Gemini 4 Carbon is out: prices, limits and how to use it

At a glanceGoogle made Gemini 4 Carbon generally available on November 10, 2026. The introductory price ends November 30. What it is, what it costs and when to pick it.

Not testedA summary of Google's announcement and documentation. The model was not run.

In this article 8 sections
  1. 1. What was announced
  2. 2. The price, and the date it changes
  3. 3. Carbon or Argon: same price list, different bill
  4. 4. The limits that matter
  5. 5. How to use it
  6. 6. What Google claims, and how to read it
  7. 7. What to know before relying on it
  8. 8. Should you switch?

On November 10, 2026, Google made Gemini 4 Carbon generally available. Two weeks earlier it had gone only to a group of security teams. Carbon is Google's strongest reasoning model, and for now it is priced the same as its lighter sibling, Gemini 4 Argon. The detail that matters for budgets is the date on the price list: the introductory rate ends on November 30, and the price doubles the next day.

1. What was announced

  • General availability. Carbon is open to paying developers and enterprises through the Gemini API, and is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity.
  • The top of the Gemini 4 line. Google describes Carbon as built for work that tests the limits of a model's reasoning: complex multi-step coding and debugging, dense documents, scientific problems and security auditing.
  • An introductory price. $2 per million input tokens and $10 per million output tokens until November 30, 2026.
  • A preview of agent teams. Google is previewing a /swarm command in Antigravity that turns one prompt into a coordinated group of agents.

2. The price, and the date it changes

Gemini 4 Carbon, until November 30$10 per million output tokens

$10

Gemini 4 Carbon, from December 1$20 per million output tokens

$20

Claude Opus 5.5, for comparison$20 per million output tokens

$20

Claude Fable 5.2, for comparison$40 per million output tokens

$40
Output price per million tokens, from each company's own price list. Bars are drawn to scale.
Per million tokensThrough November 30, 2026From December 1, 2026
Input$2.00$4.00
Output, including thinking tokens$10.00$20.00
Cached input$0.10$0.20

Three notes from Google's pricing page:

  • There is no free tier for Carbon. Every token is billed, including the model's thinking tokens.
  • Caching is automatic only. The page says explicit caching is not supported.
  • Grounding with Google Search is free for 5,000 requests a month, shared across recent Gemini models, then $14 per 1,000 requests.
Budget at the December priceAnything you build this month will run at twice the token price from December 1. Do your sums with $4 and $20, and treat November as a discount.

3. Carbon or Argon: same price list, different bill

Carbon and Argon share the same context window, the same built-in tools and the same price per token. Google's documentation says the main difference is how many thinking tokens each spends per turn. Carbon spends more, to explore and to verify its work, which raises both the token count and the wait.

Gemini 4 ArgonGemini 4 Carbon
Google's descriptionEfficient; the choice for most software and knowledge workStrongest reasoning; for tasks at the limit
Price per tokenIdentical
Tokens used per turnFewerMore
SpeedFaster repliesSlower
Google's examplesInteractive coding, day-to-day feature workRoot-cause debugging, architecture, dense multimodal analysis, security audits

So an identical price list does not mean an identical bill. A task that takes Argon 30,000 output tokens might take Carbon more, and you pay for the difference. Here is the arithmetic for one request with 200,000 input tokens and 30,000 output tokens, which are invented numbers:

Until November 30From December 1
Input, 200,000 tokens$0.40$0.80
Output and thinking, 30,000 tokens$0.30$0.60
Total$0.70$1.40

4. The limits that matter

PropertyValue
Model codegemini-4-carbon
Context window1,048,576 tokens, shared by input, thinking and output
Output in a normal requestUp to 65,536 tokens
Output with long decodeUp to the full window, minus your input, as a background job
InputsText, images, video, audio and PDF
Thinking levelslow, medium (the default), high

The shared window is worth a second look. Input, thinking and output draw on one budget, so a very large prompt leaves less room for the answer.

5. How to use it

In the Gemini app: Carbon is rolling out to Google AI Ultra subscribers. If you are on another plan, it is not listed for you yet.

Through the API: you need the paid tier. Google's quickstart, in Python:

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-4-carbon",
    input="Explain what a race condition is and show a minimal example in Go.",
)
print(interaction.output_text)

The structure is Google's; the prompt is ours. You can also try the model in Google AI Studio without writing code.

Three settings from Google's guide decide what you pay and how long you wait:

  • Thinking level. The guide calls it the main control over quality, latency and token use. Its advice is to run your own evaluations at all three levels and pick the lowest that meets your quality bar. Keep medium unless a workload needs high.
  • Output limit. max_output_tokens caps thinking and answer together. If an existing application sets it below 32,768, the guide says to raise it to at least that, because Carbon can spend more of the limit on thinking. To use fewer tokens, lower the thinking level instead.
  • Timeouts. Longer reasoning means longer requests. The migration notes say to make sure your HTTP timeouts allow for it.

6. What Google claims, and how to read it

Google's announcement says Carbon ranks first on three third-party benchmarks, CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than two times lower cost per task than the next best external model. It also reports strong results on security tasks for the version offered to vetted defenders.

Two cautions. The cost comparison was made at the introductory price, which doubles on December 1. And a benchmark tells you how a model did on someone else's tasks. The only number that should decide your choice is how it does on yours.

7. What to know before relying on it

  • Data use. On the paid tier, the pricing page states that your content is not used to improve Google's products.
  • Security features are gated. The strongest cybersecurity capabilities are offered through a separate program for vetted defenders, which is where Carbon first appeared.
  • It is the expensive choice by design. Google itself says Argon remains the right model for the majority of software and knowledge work.

8. Should you switch?

Your situationSuggestion
You use Argon and it does the jobStay. Same price per token, fewer tokens, faster.
Argon gives up or gets it wrong on your hardest tasksTry Carbon on those tasks only, and compare the result and the token count.
You want to evaluate it cheaplyDo it before November 30, while it costs half.
You are choosing between vendorsFrom December 1 its price matches Claude Opus 5.5. Compare on your own tasks, not on price.
You are on a free planNot available. Argon is also paid-only in the API.

For the other recent price change at the top of the market, see the Claude Fable 5.2 launch explainer. All release explainers are listed on the AI releases page.

We also covered this release on our channel, AI2Now: Gemini 4 Carbon is over 2x cheaper per task.

Sources

This article reflects public information available when it was written. Versions, defaults and prices can change; check the linked official documentation.

Prices, limits and availability are from Google's announcement posts and Gemini API documentation as read on November 12, 2026. This is a summary of published information: the model was not run for this article, and Google's benchmark results are reported as Google's claims. The cost example uses invented token counts to show the arithmetic. Prices change on December 1, 2026; check the pricing page before budgeting. Written for this site with AI assistance, from the sources above. For errors or additional information, please request a correction. Editorial policy