MiniMax M3.1 Flash launches: pricing, access and the cost of a finished task

What is available now?

MiniMax says M3.1 Flash is available through its API, M Plan and MiniMax Code. The launch announcement says downloadable weights will follow later; it does not give a release date. Hosted access should not be confused with being ready to run the model locally.

A product announcement establishes what the provider is offering. It does not establish how well your own backlog will run. Keep the buying decision tied to a small, measured trial.

Public record: MiniMax launch announcement — October 19, 2026

What does the API cost?

The launch post lists US-dollar prices per million tokens: $0.15 for input, $0.60 for output and $0.015 for cached input.

For illustration, one million uncached input tokens plus 200,000 billed output tokens totals $0.27 at those rates. That calculation excludes external tools and infrastructure. It also assumes those are the actual billed token volumes, rather than merely the final text a user sees.

Public record: MiniMax launch announcement — October 19, 2026

A large context window still needs an effort setting

MiniMax’s current API guide lists the model identifier as MiniMax-M3.1-Flash, with a one-million-token context window and support for text, image and video input. It describes both Anthropic-compatible and OpenAI-compatible request formats.

The guide also makes a consequential detail explicit: thinking depth can be set to low, medium, high, xhigh or max, and defaults to max. Higher settings can increase latency and output-token usage. Thinking cannot be switched off; the documented way to reduce that overhead is to lower the effort level.

For a fair comparison, keep the effort setting visible in the trial record. Comparing one model at its deepest reasoning setting with another at a faster default can obscure which configuration actually fits the job.

Public record: MiniMax API guide — supported models, context and thinking controls

Vercel offers a short testing window

In an October 19 changelog, Vercel says M3.1 Flash is free through AI Gateway until October 22. The offer applies only when MiniMax serves the request, and Vercel directs users to its provider-order configuration to ensure that routing. It says pinned requests return to MiniMax’s standard rate after the offer.

Gateway users should use the gateway-specific identifier minimax/minimax-m3.1-flash. Check the live offer and routing terms before running a trial; temporary access is not a permanent production budget.

Public record: Vercel AI Gateway availability and temporary offer — October 19, 2026

Measure a completed task before switching

Choose ten small, repeatable jobs from your own backlog. For a coding team, useful examples include repairing a known bug, adding a form validation rule and writing a component test. For operations, try reconciling two sample records or turning a mock enquiry into a structured follow-up draft. Use test data and a reviewable destination.

Write the acceptance criteria first. A bug fix should pass the relevant test without breaking another path. A follow-up draft should preserve the facts supplied and leave missing information visibly unresolved. Run each job with the same starting material in your current setup and the new model.

Record completion, retries, billed usage, elapsed time and the minutes someone spends correcting the result. Calculate total trial spend divided by accepted completions, and keep human review time beside that number. A low token bill with repeated rework may be the more expensive workflow.

Keep the existing setup available during the pilot. Expand only after the new configuration meets the quality threshold on work that resembles your actual queue. Promptwire has not independently benchmarked this release; this testing method is editorial guidance.