MiniMax

Direct access versus this gateway

The mainland OpenAI-compatible base is https://api.minimaxi.com/v1; Anthropic compatibility uses https://api.minimaxi.com/anthropic, both with a MiniMax key. This gateway uses its own base URL and Bearer token and declares only /v1/chat/completions; provider Responses, Anthropic, and media APIs are not gateway routes.

Keys, models, and thinking

Pay-as-you-go and Token Plan subscriptions use separate keys and cannot be mixed. Current text documentation centers on MiniMax-M3/M2.x, while abab remains mainly a legacy search alias. M3 can disable or adapt thinking, whereas M2.x cannot disable it; match the exact model, key type, and parameters to the current page.

Multi-turn tools and Responses

New integrations use tools; the old function_call shape is deprecated. Preserve the assistant's complete tool_calls and return tool results with the same tool_call_id. M3/M2.x reasoning may appear in content or split fields, and must remain in later turns even with reasoning_split=true. Official Responses supports SSE and tools, but gateway clients use only the paths listed here.

Limits, errors, and cost

Limits can be measured by RPM, TPM, and connections and are shared across primary and subaccounts. Code 1002 commonly covers rate limits, 1008 insufficient balance, and 1041 connection limits; retain trace_id for support. The official guide bills service_tier=priority at 1.5 times the normal rate, so enable it only intentionally.

Media and production safety

Speech, video, and music use separate tasks, billing units, rights, and moderation rules rather than text Chat. Cap output, concurrency, polling, and retries, and alert on balance. Move generated media into controlled storage and verify voice, likeness, and music rights.