MiniMax M3.1 Flash launches: Cheap coding, but no model weights yet

MiniMax launched M3.1 Flash on October 19, 2026. It is available through the API, M Plan subscription and MiniMax Code. The company calls it an open-weight model, yet says in the same announcement that the weights will be released later. It would be misleading to present it today as a downloadable model you can already run on your own server.[1]

This is for a developer building a budget coding agent or a long-document app who needs to decide whether to pilot M3.1 Flash against the current model. My short answer: the price and independent quality data justify a trial, but speed and cost per completed task need verification on your own workload.[1][2][5]

Where does it stand on coding and agent tests?

MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on the OSWorld V2 computer-use test and 29.6 on Agents' Last Exam. It also claims agent-task usage costs 70% less than the previous M3 and online inference is nearly three times faster. These are vendor measurements. I have not run M3.1 on the same tasks, and those percentages are not a promise about your own time or bill.[1]

Artificial Analysis independently reports a 43 on its Intelligence Index, an average $0.22 per task in that test and 116 output tokens per second. Its page calls the model strong for the price but slightly slower than average and very verbose: 180 million output tokens over the evaluation, versus a 99-million median. Do not treat that 116 and MiniMax's claimed 170+ median tokens per second as the same measurement; workloads and methods differ. Verbose output also changes what a low token rate means on your invoice.[1][5]

What work should you test?

The API documentation lists a one-million-token context, text/image/video input, tool use and adjustable reasoning effort. Effort supports low, medium, high, xhigh and max; omitted effort defaults to max, and thinking cannot be switched off. Do not run a simple file edit at the default max and assume the token bill represents the cheapest usable setting. Try low and medium against the same acceptance test first, then increase for hard tasks.[3]

In MiniMax's own 200-task ProgramBench experiment for Codex, a 512,000-token window performed similarly to the full million-token window while reducing estimated usage-based charges by 19.9%. That is a vendor test, not a guaranteed saving for your project. For a codebase or document review, use the larger window only when the task actually needs it.[7]

API pricing is not the same as a subscription

The MiniMax API rate card lists M3.1 Flash at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. The old M3's standard API rates at or below 512,000 input tokens are $0.30/$1.20; Flash halves those unit prices. The company's 70% lower agent-task-cost claim measures something else. A 50% cut in token rates guarantees neither 50% nor 70% savings on every task.[1][2]

M Plan's Go, Explore and Build tiers all include Flash. You can sign into MiniMax Code with the subscription or connect a plan key to other tools; pay-as-you-go production API traffic is a separate billing choice. The plan describes tier allowances relative to Go, but those ratios should not be presented as a raw token entitlement. Keep product surface and bill separate.[6]

Who should switch, and who should wait?

For price-sensitive coding or document jobs with testable outputs, I would run a three-task pilot: a small bug fix, a multi-file change and a tool-using summary. Run the same checks on your current model and M3.1, then record total tokens, elapsed time, accepted output and human correction rounds. Since the model can be verbose, measure completion rather than only answer speed.[3][5]

Wait if self-hosted weights in your own data center are mandatory: MiniMax has not published them yet. Current M2.5 and M2.7 users also have a migration deadline. New-user access ended October 19; existing pay-as-you-go API users can continue until November 18, 2026. M Plan and Token Plan requests have been routed to M3.1 Flash since October 19. An old model ID still working may not mean the old model is still serving the request.[1][4]

The decision

M3.1 Flash is a serious budget candidate for coding and agents. Independent testing paints a quality-strong but verbose picture and offers a more cautious speed reading than the vendor launch. Those who need local weights should wait. API and subscription users should compare acceptance tests and cost per completed job before switching. I have not tested the model hands-on; this is a pilot recommendation grounded in the published evidence.[1][2][5]