Gemini 4 Carbon API pricing and prompting guide

Price per million tokens and specifications

On KeelRouter, Gemini 4 Carbon is $1 per 1M input tokens and $5.20 per 1M output tokens, with cached input reads at $0.10 per 1M tokens. Thinking tokens are billed as output.

Specifications: 1,048,576-token context window, up to 65,536 output tokens per response, and multimodal inputs (text, image, audio, video, and PDF). On Artificial Analysis, Carbon scores 60 on the Intelligence Index, finishes an average index task in 26.1 minutes (66 seconds to first answer token), and ranks #1 on AutomationBench-AA (77.8%) and Terminal-Bench 4.0 (65.2%).

USD per 1M tokens, checked 2026-11-12. Google's Gemini 4 Carbon uses implicit prefix caching; always set max_tokens on open-ended tasks so deep reasoning loops stay within budget.

When to use Carbon vs Gemini 4 Argon

Google positions gemini-4-argon ($1 / $5.20 per 1M tokens on KeelRouter) as the default workhorse for everyday coding and knowledge work, and gemini-4-carbon as the escalation model when a problem needs deeper exploration and self-verification. Although both have the same per-token price on KeelRouter, Carbon spends more thinking tokens exploring hypotheses and checking its own work before answering.

Use Carbon when: (1) a multi-file bug or race condition resisted Argon or Sonnet 5.5; (2) an agent must coordinate shell commands, browser actions, and structured files across many steps (where Carbon leads Terminal-Bench 4.0 at 65.2% and AutomationBench-AA at 77.8%); or (3) you need deep verification over long PDFs, video, or audio transcripts (where Google reports 92% on its agentic honesty benchmark vs 70% for Argon).

Prompt templates for Gemini 4 Carbon

Carbon performs best when you give it the raw evidence (logs, diffs, PDFs, or command outputs), state the verification criteria upfront, and ask it to falsify its own initial hypothesis before proposing a fix.

You are debugging a production failure across the files and logs below.
1. List the top 3 plausible root causes that explain every symptom in the trace.
2. For each hypothesis, cite the exact lines or log timestamps that support or contradict it, and eliminate any hypothesis that requires an unverified assumption.
3. Produce the minimal patch for the surviving root cause, then trace how the patched code behaves under concurrent retries, timeouts, and malformed input.

Execute this environment setup and migration task end to end.
Rules:
- After every shell command or file edit, verify the exit code and inspect the actual output before deciding the next command.
- Never assume a service is running, a port is free, or a schema column exists without checking.
- If a command fails twice with the same error, inspect the logs and environment variables instead of re-running the same flag.
- End with a machine-readable JSON report: {"status": "completed" | "blocked", "verified_checks": [...], "changed_files": [...]}.

Audit the attached specification and financial tables for internal contradictions, missing edge cases, and unit mismatches.
Hard constraint: every claim in your report must quote the exact section and figure from the source text. If a number cannot be derived from the provided pages, mark it "UNVERIFIED IN SOURCE" rather than estimating.

curl https://api.keelrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $KEELROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-4-carbon",
    "max_tokens": 16384,
    "messages": [
      {
        "role": "user",
        "content": "Find the race condition in this async queue consumer and prove why it can double-settle a job."
      }
    ]
  }'

FAQ

How much does the Gemini 4 Carbon API cost on KeelRouter?

$1 per 1M input tokens and $5.20 per 1M output tokens (including thinking tokens), with cached input reads at $0.10 per 1M tokens. Google's official list price is $2 / $10 through November 30, 2026 and $4 / $20 from December 1.

Why does Gemini 4 Carbon use more tokens per call than Argon?

Carbon is built to explore multiple hypotheses and run internal verification loops before answering, and those thinking tokens bill at the output rate. Cap max_tokens on open-ended requests, and use Gemini 4 Argon for routine tasks.

How does Gemini 4 Carbon compare with Claude Opus 5.5 and Fable 5.2?

On Artificial Analysis, Gemini 4 Carbon scores 60 on the Intelligence Index—ahead of Claude Opus 5.5 at max effort (59) and behind Claude Fable 5.2 at high/max (62–65)—while leading both on Terminal-Bench 4.0 (65.2%) and AutomationBench-AA (77.8%) at a lower cost per task.

Can I call Gemini 4 Carbon and Claude with the same API key?

Yes. One KeelRouter API key and one USD balance call gemini-4-carbon, gemini-4-argon, claude-opus-5-5, and claude-fable-5-2 on /v1/chat/completions.

More guides

• GPT Image 2.6 API pricing
• GPT Image 2.6 prompts
• Make a photo sing with an API
• AI face swap with lip sync: 8 ways we tried on the White Chicks car scene
• Claude Fable 5.2 API pricing and how to call it
• Claude Fable 5.2 prompting guide and community templates