# MiniMax M3

`minimax/minimax-m3` · MiniMax · text

M3 can reason over up to one million tokens in one request, enough to compare lengthy specifications or trace dependencies across a large codebase.

[Model page](https://sociaro.com/models/minimax/minimax-m3) · [JSON](https://sociaro.com/models/minimax/minimax-m3.json) · [Model pricing](https://sociaro.com/pricing?model=minimax%2Fminimax-m3) · [Documentation](https://docs.sociaro.com/models/minimax-minimax-m3/)

## Use cases

- Trace a behavior across several parts of a codebase
- Compare requirements across lengthy specifications

## Technical details

- API model ID — minimax/minimax-m3
- Provider — MiniMax
- Modality — Text
- Input — Chat messages
- Output — Text
- Endpoint — POST /v1/chat/completions

## Request parameters

- `messages` (object[], required) — The conversation, in OpenAI's format.
  - OpenAI chat messages (system, user, assistant, tool)
- `max_completion_tokens` (integer, optional, default `the model's`) — Thinking tokens count toward it: too small a cap ends with finish_reason: length and an empty answer.
  - 1 up to the model's output limit
- `max_tokens` (integer, optional) — The older name for the same cap; MiniMax marks it deprecated.
  - as above
- `stream` (boolean, optional, default `false`) — Server-sent events, as in OpenAI's API.
  - true · false
- `stream_options` (object, optional) — Adds a final chunk with token usage.
  - {"include_usage": true}
- `temperature` (number, optional, default `1`) — MiniMax documents it as having no effect on M3 (fixed at 1).
  - (0, 1]
- `top_p` (number, optional, default `0.95`) — MiniMax documents it as having no effect on M3 (fixed at 0.95).
  - (0, 1]
- `tools` (object[], optional) — Function calling, in OpenAI's format.
  - OpenAI function tools
- `tool_choice` (string, optional, default `auto`) — Whether the model may, must, or must not call a tool, as MiniMax documents for M3.
  - none · auto · required
- `thinking` (object, optional, default `on`) — The model reasons before answering by default; disabled turns it off.
  - {"type": "adaptive"} · {"type": "disabled"}
- `reasoning_split` (boolean, optional, default `the vendor's`) — true returns the reasoning in reasoning_content; false leaves it inside the answer between think tags.
  - true · false
- `service_tier` (string, optional, default `standard`) — MiniMax's priority tier, on M3 only, billed at 1.5× the standard rates.
  - priority

## Pricing

Rates in USD. Catalogue timestamp 2026-11-21T02:23:44Z.

- USD 0.36 per 1M input tokens · active
- USD 0.72 per 1M input tokens · prompts over 512,000 tokens · active
- USD 0.54 per 1M input tokens · priority (service_tier: priority) · active
- USD 1.44 per 1M output tokens · active
- USD 2.88 per 1M output tokens · prompts over 512,000 tokens · active
- USD 2.16 per 1M output tokens · priority (service_tier: priority) · active
- USD 0.072 per 1M cached input tokens · active
- USD 0.144 per 1M cached input tokens · prompts over 512,000 tokens · active
- USD 0.108 per 1M cached input tokens · priority (service_tier: priority) · active
- USD 0 per 1M cache-write tokens · active · free
- USD 0 per 1M cache-write tokens · priority (service_tier: priority) · active · free

- A prompt over 512,000 tokens bills the whole request at the long-context rates.
