MiniMax M3.1 Flash: API Pricing and Always-On Reasoning

MiniMax M3.1 Flash launched on October 19 with API pricing of $0.15 per million input tokens and $0.60 per million output tokens. It supports a million-token context window and always-on reasoning. Downloadable weights are still forthcoming, with MiniMax announcing an October 23 release plan.

The launch announcement identifies Space Bunny as an earlier anonymous version of the model. M3.1 Flash is available through the MiniMax API, M Plan and MiniMax Code. For developers, the immediate questions are how much a complete agent task costs and whether an existing integration handles the new reasoning behavior correctly.

What MiniMax M3.1 Flash costs

MiniMax's API pricing page lists three rates per million tokens: $0.15 for input, $0.60 for output and $0.015 for cached-input reads. The announcement describes these as flat rates, and the published Flash row has no separate long-context tier.

The input and output rates are half the standard M3 rates for requests with up to 512,000 input tokens. M3's higher-context tier costs more again. This matters when reading MiniMax's separate claim that overall agentic-task usage costs fell by 70%. A task-level reduction can reflect how many tokens a workflow uses; it is not the same comparison as a reduction in the price of each token.

As a simple API-only example, a batch of requests totaling one million uncached input tokens plus 100,000 output tokens would cost $0.21 at the published Flash rates. That arithmetic assumes those are the billed token counts. It excludes tool charges, infrastructure and any extra calls an agent makes while retrying or checking its work.

M Plan access is a separate subscription route. Its inclusion in the launch does not mean a subscription has the same allowances as pay-as-you-go API billing.

Reasoning cannot be switched off

The main migration issue is in the model invocation documentation. The API model identifier is MiniMax-M3.1-Flash, and reasoning effort has five settings: low, medium, high, xhigh and max. Omitting the setting selects max.

An integration that sends thinking.type as disabled, or effort as none, receives an HTTP 400 error. To request less reasoning, use a lower supported effort level. Reusing a configuration that disabled thinking for another model can therefore break a request before the agent does any work.

The field also depends on the protocol. Anthropic-compatible requests use output_config.effort; Chat Completions uses reasoning_effort; Responses uses reasoning.effort. On the OpenAI-compatible Chat Completions route, reasoning_content holds the thinking output and content holds the answer.

For multi-step tools, check the Anthropic-compatible API reference before trimming conversation history. Returned thinking blocks need to be preserved for subsequent tool-use turns. A wrapper that works for a single text response may still mishandle a longer exchange.

These details make effort a deployment setting worth testing explicitly. Run the same small set of representative tasks at a lower setting and at max, then compare completed work, billed tokens and elapsed time. Leaving the default untouched makes it harder to tell whether extra reasoning earns its cost on your workload.

Context size, speed and benchmark claims

The million-token window gives an agent room for substantial source material and a long conversation. It does not establish that every detail will be used reliably. A useful repository test should ask for evidence from distant files and check whether later tool calls still respect earlier constraints.

MiniMax reports 73.7% on DeepSWE v1.1 and 61.7 points on OSWorld V2. It also reports median output speeds above 170 tokens per second, with speeds exceeding 200 possible. These are the publisher's results, rather than measurements from this article, and output speed alone does not describe time to finish a task involving tools, retries and review.

The documented rate-limit table lists 200 requests per minute and 10 million tokens per minute for M3.1 Flash. Those are published interface limits, not a guarantee that every workload will sustain that throughput. Check the limits applying to the account and access route before sizing a service around them.

Hosted access now, model weights later

MiniMax's official launch post says model weights will arrive on October 23. That is an announced plan, not a downloadable release at publication. Teams considering self-hosting still need the actual artifacts, license and deployment requirements before estimating hardware or making a rollout commitment.

For API users, the useful next step is more immediate: verify the reasoning settings, test a complete tool loop and measure cost per accepted result. The lower token rates make experimentation less expensive. Whether they lower the cost of useful work depends on how often the model completes the task without another attempt or a substantial correction.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI and developer-tool teams turn model capabilities, integration choices and real product tradeoffs into clear technical stories that reach the right users.

Book a 30-minute call.