Gemini 4 Carbon — reasoning depth for your hardest problems
Gemini 4 Carbon is Google's strongest Gemini 4 model. It spends more thinking per turn than Gemini 4 Argon to explore, verify and check edge cases before it answers. $0.90 / 1M input · $4.50 / 1M output.
- 1M-token context
- Long decode up to 1M
- Three thinking levels
- Function calling
- Multimodal input
Playground
Send native Gemini 4 Carbon requests with your account balance. Select an available group and inspect the request before you send it.
gemini-4-carbon$0.90 in · $4.50 out per 1M tokensMore Gemini parameters
Try one of these prompts
Messages are billed from your balance at the rates below, like an API call. Failed requests are free.
Gemini 4 Carbon pricing
Gemini 4 Carbon rates below come from current billing settings for each account group. Thinking counts as output, so a hard prompt costs more than its visible answer suggests.
Example sizes are for illustration. Your bill follows the tokens each request reports. Standard processing rates, with no context-length price step. Cached input is billed once at the cached rate. Search and Maps grounding are billed separately from these token rates.
- Input
- $0.90Price · 1M tokens
- Output, including thinking
- $4.50Price · 1M tokens
- Cached input
- $0.045Price · 1M tokens
What is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind's frontier reasoning model for long, multi-step work: complex debugging, architecture reviews, dense document analysis, proofs and security audits. Google made it available to paying Gemini API developers on November 10, 2026.
Gemini 4 Carbon shares Gemini 4 Argon's request format, 1M-token context and built-in tools. The difference is effort: Gemini 4 Carbon uses more thinking tokens per turn, which adds latency and usage in exchange for deeper checking.
Official model page and migration guide checked November 11, 2026. Input, thinking and output share one context window.
Start with Gemini 4 Carbon
Pick one hard task you can grade with Gemini 4 Carbon, then keep the prompt, settings and answer together.
Open the Playground
Send Gemini 4 Carbon a real problem with its supporting files. Inspect the native request JSON before you send it.
Select an available group
Choose among the Gemini 4 Carbon groups on your account. Token rates update for the group you select.
Connect your application
Copy the Gemini 4 Carbon examples from the API reference. Keep gemini-4-carbon in the endpoint path when you move the request to your server.
What Gemini 4 Carbon is good at
Use it where a wrong answer costs more than a slow one.
Find the real root cause
Give Gemini 4 Carbon the failing test, the stack trace and the code paths involved. Ask it to isolate the cause and check the fix against your tests before you merge.
Read dense documents
Gemini 4 Carbon reads long PDFs, charts and tables in one request. Ask for answers tied to the source page, and a structured output that follows your schema.
Review code for security
Ask Gemini 4 Carbon to trace how a flaw could be reached and to propose a root-cause patch. Treat every finding as a lead your team confirms.
Write very long outputs
With long decode, Gemini 4 Carbon can keep going past 64K tokens, up to 1M minus your input, by continuing from where the last response stopped.
Practical Gemini 4 Carbon tasks
Tasks worth testing; no benchmark or customer outcome is claimed for SeedRouter.
Plan a multi-package migration
Share the dependency graph and failing builds. Gemini 4 Carbon can order the steps and flag the packages that break first.
Check a proof or derivation
Paste a derivation and ask Gemini 4 Carbon to verify each step, marking any gap it cannot close.
Audit a financial or legal file set
Load several reports into one Gemini 4 Carbon request and ask which claims each source supports, with page references.
How to use Gemini 4 Carbon
- 01
Prepare your account
Sign in, add credit and create an API key for server-side use. Open Gemini 4 Carbon to see your groups and rates.
- 02
Assemble the conversation
Put user messages, earlier model parts and tool results in contents. Give Gemini 4 Carbon the files it needs with correct MIME types.
- 03
Set thinking and output
Keep medium thinking for most Gemini 4 Carbon work and use high for the hardest problems. Google recommends at least 32,768 output tokens.
- 04
Handle the result
Read Gemini 4 Carbon candidates, finishReason and usageMetadata. Keep full content parts and thought signatures when the conversation continues.
The native Gemini 4 Carbon API
Gemini 4 Carbon uses the native generateContent and streamGenerateContent endpoints. Switching from Gemini 4 Argon needs no request changes besides the model ID.
Use the stable identifier
The URL contains gemini-4-carbon. Gemini 4 Carbon needs no preview suffix and no separate name for streaming.
Leave sampling out
Gemini 4 Carbon ignores temperature, topP and topK, and explicit context caching is not supported. Implicit caching applies on its own from 4,096 prompt tokens.
Continue long outputs
A response that ends with finishReason CONTINUATION carries a continuationToken. Send the same Gemini 4 Carbon request again with that token to get the next part.
Gemini 4 Carbon vs Gemini 4 Argon
Google positions the two Gemini 4 models by thinking effort, not by API features. Gemini 4 Carbon fits hard, high-stakes tasks; Gemini 4 Argon fits fast everyday loops.
| Setting | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Thinking per turn | More tokens, slower answers | Fewer tokens, faster answers |
| Google's suggested use | Root-cause debugging, proofs, security audits | Interactive coding, day-to-day feature work |
| Default thinking | medium | medium |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Request format | Same as Argon | Same as Carbon |
Run the same task on both with explicit thinking and output limits. SeedRouter has not measured a quality or latency winner. Google Gemini 4 Carbon guide
Gemini 4 Carbon limits
One shared window
Input, thinking and output share 1,048,576 tokens. A single Gemini 4 Carbon response stops at 65,536 tokens when streaming, or 32,768 without streaming, unless you continue it.
Text out, media in
Gemini 4 Carbon reads images, audio, video and PDFs and returns text. It does not generate images or audio, and it has no Live API.
Expect longer runs
Hard prompts can take several minutes. Use streaming for long tasks, and lower the thinking level when latency matters more than depth.
Verify advanced tools
Gemini 4 Carbon lists Search, Maps, code execution, URL context, file search and computer use. Check each tool on your account group before you rely on it.
Use Gemini 4 Carbon with SeedRouter
One balance across models
Pay for Gemini 4 Carbon from the same SeedRouter balance as the other listed models. Credit never expires.
Current rates in context
See Gemini 4 Carbon prices beside the model settings. The Playground estimates token cost with your selected group's rate.
Failed requests are free
Failed Gemini 4 Carbon requests are not charged. On success, usage shows input, cached input, output and thinking separately.
Explore other text models
Compare available text models by their own documented capabilities and current rates.
Call GPT-6 Astra with the OpenAI SDK. OpenAI's most capable model, with a 1.05M-token context, live per-token pricing and a playground you can try right now.
View pricingCall GPT-6 Sol with the OpenAI SDK. 1.05M-token context, reasoning effort from none to max, live per-token pricing and a playground you can try right now.
View pricingCall GPT-6.1 Sol with the OpenAI SDK: near-Astra coding and agent work at a fifth of the price, 1.05M-token context, 95% off cached input. Try it now.
View pricingCall GPT-6 Luna with the OpenAI SDK. The lowest-cost GPT-6 model, with a 1.05M-token context, live per-token pricing and a playground you can try right now.
View pricingTry Claude Fable 5.2 for coding and problem solving with a 1M-token context. Compare live token prices and model tiers, then test it in the playground.
View pricingTry Claude Mythos 5 for complex reasoning and research with a 1M-token context. Compare live token prices and model tiers, then test it in the playground.
View pricingGemini 4 Carbon guides
Hands-on notes on Gemini 4 Carbon settings, costs and when to pick it over Argon.
Gemini 4 Carbon opened to paying API developers on November 10, 2026. Where you can use it, the price change on December 1, and what an API call returned.
Gemini 4 Carbon and Argon share one request format and price. We ran the same three prompts on both through the API: thinking tokens, latency and answers.
Gemini 4 Carbon FAQ
What is Gemini 4 Carbon?+
Gemini 4 Carbon is Google DeepMind's strongest Gemini 4 reasoning model, released to paying developers on November 10, 2026. It is built for long agentic work, complex coding, scientific reasoning and cybersecurity analysis.
Which Gemini 4 Carbon model ID should I use?+
Use the stable ID gemini-4-carbon in the endpoint path. Set the thinking level in generationConfig rather than adding a suffix to the model name.
Should I use Gemini 4 Carbon or Gemini 4 Argon?+
Start with Argon for interactive coding and fast loops. Move a task to Gemini 4 Carbon when Argon's answers miss edge cases, since Carbon spends more thinking tokens to check its work.
What thinking level does Gemini 4 Carbon use by default?+
Gemini 4 Carbon defaults to medium and also supports low and high. Google suggests testing all three and keeping the lowest level that meets your quality bar.
How long can a Gemini 4 Carbon answer be?+
One Gemini 4 Carbon response returns up to 65,536 tokens when streaming. With continuation requests, total output can reach 1,048,576 tokens minus your input.
How is Gemini 4 Carbon billed?+
Gemini 4 Carbon bills input, cached input and output tokens, and thinking counts as output. Google's introductory price runs through November 30, 2026; this page always shows SeedRouter's current rate.
Can I choose a Gemini 4 Carbon account group?+
Select a Gemini 4 Carbon group in the Playground, or send X-SeedRouter-Group with an API request. Leave the header out for automatic selection.
Try Gemini 4 Carbon
Bring the problem your current model keeps getting wrong. Run Gemini 4 Carbon on it, read the usage, and decide on your own results.