MiniMax M3.1-Flash-Preview arrives in Code and Token Plan, not general pay-as-you-go
By AgentRiot Editorial
The new model ID adds adjustable reasoning effort to a 1M-context, multimodal coding model. MiniMax is keeping this preview inside its coding app and subscription plan for now.
MiniMax has introduced MiniMax-M3.1-Flash-Preview in MiniMax Code (opens in new tab) and made it available to existing Token Plan subscribers (opens in new tab). The September 27 posts pitch it as a faster, lighter option for development and high-volume work. The company's model guide (opens in new tab) adds an important boundary: M3.1-Flash-Preview is available only through Token Plan and MiniMax Code for now. A model appearing in the API documentation is not the same as a broadly priced pay-as-you-go launch.
The useful change is adjustable thinking
MiniMax lists a 1,000,000-token context window and text, image and video input for M3.1-Flash-Preview. Those are substantial capabilities, but MiniMax-M3 already has both (opens in new tab). This preview's more specific developer-facing change is control over thinking depth: low, medium, high, xhigh or max. The model invocation guide (opens in new tab) says max is the default when the field is omitted. Higher settings use more thinking tokens and add latency, so an agent that makes frequent tool calls may need a different setting from one working through a long, difficult task.
Thinking itself cannot be turned off in this preview. MiniMax says a disabled-thinking request or none effort returns HTTP 400. On its Anthropic-compatible interface, the field is output_config.effort and thinking arrives in a separate thinking block. On the OpenAI-compatible Chat Completions interface, use reasoning_effort; the reasoning is returned in reasoning_content, separate from the answer in content. The OpenAI-compatible API reference (opens in new tab) also says a multi-turn tool-calling app must retain the complete assistant response, including tool calls and reasoning content, to preserve continuity. That detail matters more to a coding-agent integration than the word “Flash” in the name.
The published model ID is exactly MiniMax-M3.1-Flash-Preview. MiniMax documents Anthropic-compatible and OpenAI-compatible request formats, but the Token Plan restriction (opens in new tab) still governs access. Teams should not assume their existing standard pay-as-you-go API key can invoke it merely because the endpoint format looks familiar.
The subscription is clear; the per-token rate is not
MiniMax says current Token Plan subscribers get M3.1-Flash-Preview without extra setup. The Token Plan pricing page (opens in new tab) lists Plus at $22/month, Max at $55/month and Ultra at $132/month, with rolling five-hour and weekly quota windows. Those are plan prices and quotas, not an M3.1 per-token tariff. MiniMax's pay-as-you-go rate card (opens in new tab) lists M3 and older language models, but does not list M3.1-Flash-Preview as of September 27. Applying M3's token price to the preview would be a guess.
The text of MiniMax's Code announcement (opens in new tab) and Token Plan announcement (opens in new tab) describes speed and reliability without a measured latency figure or coding-benchmark score for M3.1. Treat “faster” as MiniMax's positioning, not an independently established performance gain. Subscribers can try the exact model ID, compare effort settings on their own agent tasks and watch token use and latency. Teams needing pay-as-you-go access or a measured speed-versus-quality comparison still need those details from MiniMax.