MiniMax M3.1 Flash in Claude Code: Setup, Config and Fixes

What you get

Claude Code talks to an Anthropic-compatible API, and MiniMax exposes exactly that shape at https://api.minimax.io/anthropic. Point Claude Code at that base URL and it runs the whole agent loop on M3.1 Flash: a 1,000,000-token context window, image and video input, and reasoning that is always on with adjustable depth.

Option 1: the one-click setup wizard

MiniMax ships a CLI that configures the agent for you:

npx -y mmx-cli@latest agent setup

Select Claude Code in the wizard. It installs the agent if the claude command is missing, checks claude --version, updates ~/.claude/settings.json and adds the MiniMax model list. The wizard also backs up the settings file it is about to change.

Option 2: edit settings.json manually

Open ~/.claude/settings.json and set the environment block:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.minimax.io/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "<MINIMAX_API_KEY>",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "524288",
    "ANTHROPIC_MODEL": "MiniMax-M3.1-Flash[1m]"
  }
}

Four details are easy to get wrong:

ANTHROPIC_BASE_URL must point at the Anthropic-compatible path, /anthropic, not /v1.
ANTHROPIC_AUTH_TOKEN holds the MiniMax API key.
CLAUDE_CODE_AUTO_COMPACT_WINDOW=524288 sets the auto-compaction threshold in tokens. M3.1 Flash accepts a 1M-token input window; the harness still compacts earlier by default, and this value changes that.
ANTHROPIC_MODEL=MiniMax-M3.1-Flash[1m] — the [1m] suffix is what requests long-context behaviour from the harness. Without it you may get a much smaller effective window.

Clear conflicting environment variables first

This is the single most common reason a correct settings.json appears to do nothing. Environment variables take priority over the settings file, so an inherited ANTHROPIC_AUTH_TOKEN pointing at a different provider silently wins.

unset ANTHROPIC_AUTH_TOKEN
unset ANTHROPIC_BASE_URL

And if those are exported in ~/.bashrc or ~/.zshrc, remove the lines — otherwise every new shell reloads them and the problem comes back.

Option 3: CC Switch

MiniMax also documents CC Switch as a third route for managing the configuration. If you prefer a GUI or already use CC Switch for other providers, follow the official page rather than hand-editing the file.

Verify it works

Start Claude Code with a short prompt and confirm the model name shown in the session header is MiniMax-M3.1-Flash.
Ask it to repeat the base URL or print a token count; a request that never leaves your machine will not produce a MiniMax response.
Send a deliberately long document and check that the context is not compacted immediately.

Troubleshooting

Symptom	Cause	Fix
Still uses Claude, not MiniMax	`ANTHROPIC_BASE_URL` or `ANTHROPIC_AUTH_TOKEN` exported in the shell	Unset both, remove the shell rc lines, open a new terminal
Model not found error	`ANTHROPIC_MODEL` still uses a preview-era id	Use `MiniMax-M3.1-Flash[1m]`
Context compacts far too early	Default auto-compact window	Set `CLAUDE_CODE_AUTO_COMPACT_WINDOW=524288`
Costs more than expected	Reasoning effort defaults to `max`	Lower `effort` in the API call path; see pricing
Requests fail on long input	Context window not applied	Keep the `[1m]` suffix in the model id

Cost note

Long Claude Code sessions are the classic way to get a surprising invoice: a 1M-token context, re-sent every turn, plus reasoning tokens that are always on. Keep the stable prefix cacheable and check the effort setting. The pricing page has the arithmetic.

The alternative path

Ollama can launch Claude Code against the same model in one command:

ollama launch claude --model minimax-m3.1-flash:cloud

That path is shorter but routes through Ollama’s cloud, with its own pricing and provider behaviour. The OpenCode and agents page compares the two routes.

Next: the Codex guide covers the same model in OpenAI’s coding agent, and the FAQ collects the errors people hit most.

Use MiniMax M3.1 Flash in Claude Code