Readme
Gemini 4 Carbon
Overview
Gemini 4 Carbon is Google’s strongest reasoning model. It is built for long, multi-step work: complex software engineering and debugging, analysis across dense documents and data, scientific and mathematical reasoning, and cybersecurity defense.
Carbon thinks longer than Gemini 4 Argon before it answers. It explores more hypotheses, checks its work, and looks for edge cases. That makes it slower and more token-hungry per request, and a better fit for the hardest problems.
Thinking levels
The thinking_level input is the main control over quality, latency, and token use.
| Level | Best for | Latency |
|---|---|---|
| low | Simpler tasks and cost-sensitive work | Fastest |
| medium (default) | Most tasks, including complex code and agentic use | Moderate |
| high | Hard math and science, and the most difficult code | Slowest |
At high, a hard prompt can take several minutes before the first token arrives.
Inputs
- prompt: Text prompt to send to the model
- images: Up to 10 images (each up to 7MB)
- videos: Up to 10 videos (each up to 45 minutes)
- video_fps: Frame rate used to sample input videos (default 1 fps)
- audio: One audio file (up to 8.4 hours)
- system_instruction: Guide the model’s behavior
- thinking_level: Control reasoning depth (low, medium, high)
- max_output_tokens: Maximum tokens to generate (up to 65,535)
max_output_tokens caps thinking tokens and answer tokens together. If you lower it, keep it at 32,768 or above so that thinking does not cut the answer short. To spend fewer tokens, lower thinking_level instead.
Gemini 4 Carbon does not use temperature or top_p, so this model has no sampling inputs.
Output
The model streams text. It has a 1 million token context window.
Pricing
Pricing is per token. Thinking tokens are billed as output tokens.
| Period | Input | Output |
|---|---|---|
| Introductory, through November 30, 2026 | $2 / 1M tokens | $10 / 1M tokens |
| From December 1, 2026 | $4 / 1M tokens | $20 / 1M tokens |