Gemini 4 Carbon API

Updated 12 Nov 2026

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google's most capable reasoning model and the second model in the Gemini 4 family, announced October 28, 2026 and opened to developers on November 10, 2026. It is built for long agentic coding runs, knowledge work in finance and law, scientific reasoning, and security defense. Input and output share one 1M-token window, and one response returns up to 64K tokens. It runs on UnificAlly as google/gemini-4-carbon, with text and image input and three reasoning effort levels (low, medium, high).

Gemini 4 Carbon key features

What Gemini 4 Carbon is best for

• Long coding agents

  63.1% on Terminal-Bench 4.0 and 89.1% on Vibe Code Bench v1.1 for multi-step builds and repairs.

• Finance analysis

  #1 on the Vals Finance Index at 67.1% for filings, models, and reconciliations.

• Tax and legal research

  #1 on Tax Agent Bench at 51.8% and Legal Research Bench at 55.8%.

• Security review

  88.2% on SEC-bench Pro and 71.0% on CWE-bench for finding a bug and writing the fix.

• Whole-codebase context

  A 1M-token window with up to 64K tokens out in one response.

• Hard multi-step reasoning

  60 on the Artificial Analysis Intelligence Index at high effort.

Gemini 4 Carbon use cases

Hand it the long jobs. Run a coding agent that works through a failing build in the terminal, reads the logs, and keeps going until the tests pass. Build finance and tax pipelines that read filings, check the numbers against each other, and draft the memo. Use it for legal research that has to find the right authority and cite it. Point it at a service for a security review that finds the vulnerability, explains the risk, and writes the patch. Load a large repository or a long document set into one request and ask questions across all of it.

Gemini 4 Carbon limitations

It is slow and it writes a lot. Artificial Analysis measures about 80 output tokens per second and more than a minute before the first answer token at high. It used 205M output tokens to finish the Intelligence Index against a median of 72.8M. Thinking tokens bill as output, so a long job costs more than the per-token rate suggests.

It does not lead every board. It ranks third on the Vals Index, and Gemini 4 Argon still beats it on Vibe Code Bench v1.1 (91.9% against 89.1%) and the Vals Legal Index (76.8% against 75.9%).

Google measured the security scores without the safety filters the public model ships with. Expect refusals on offensive security requests.

Images go in as direct https links or inline base64 data URIs, and a link that redirects is refused. Video and audio input are not available on this model yet, and one response tops out at 64K tokens.

Gemini 4 Carbon vs Gemini 4 Argon

Carbon is the stronger reasoner. The Artificial Analysis Intelligence Index goes from 50 to 60, the Vals Index from 69.4% to 70.5%, Terminal-Bench 4.0 from 57.6% to 63.1%, and Tax Agent Bench from 49.2% to 51.8%. Argon answers sooner, spends about half the output tokens on the same Intelligence Index run, and keeps the lead on Vibe Code Bench v1.1. Both share the 1M-token window and the same three effort levels.

When to use Gemini 4 Carbon

Use Gemini 4 Carbon when the task is long, the answer has to be right, and waiting is fine: terminal agents, finance and tax analysis, legal research, and security review. Stay on Gemini 4 Argon for everyday coding and knowledge work, where it is quicker and uses fewer tokens. Drop to Gemini 3.8 Flash when speed matters more than depth.

Gemini 4 Carbon API examples

Call Gemini 4 Carbon from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Full parameter docs live at docs.unifically.com/models/llm/google/gemini-4-carbon.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "google/gemini-4-carbon",
    "reasoning_effort": "high",
    "stream": true,
    "messages": [
      { "role": "user", "content": "Review this module for security bugs, rank them by severity, and write a patch for the worst one." }
    ]
  }'

Tokens arrive as they generate. Streaming is the safer default on this model, since a reply at high can take minutes. Lower reasoning_effort to low for quick steps.