MiniMax launched M3.1 Flash on October 19, 2026, publishing low token rates for developers running coding and other tool-using agents. MiniMax says M3.1 Flash is available through its API, M Plan subscription and MiniMax Code, with model weights to follow.
The pricing gives teams a concrete basis for estimating API spend. Its reasoning settings deserve attention too: the current API documentation requires reasoning and defaults to maximum effort. MrKeyoor reviewed the announcement and documentation for this report. We haven't run the model or verified MiniMax's performance claims.
What the listed rates buy
MiniMax lists $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens in its pricing table. For illustration, one million uncached input tokens plus one million output tokens would cost $0.75 at those rates. That's a calculation from the tariff, without tools or other charges, rather than a measured agent run.
In its announcement, MiniMax claims that overall usage cost for agentic tasks is 70% lower than its predecessor. It also reports 73.7% on DeepSWE v1.1, a benchmark of software tasks involving implementation and testing. Both figures are vendor results. Neither establishes how often the model will finish your team's work correctly on the first attempt.
A useful evaluation would keep the task and acceptance checks fixed, then record total tokens and retries. That would turn the published token prices into a cost per accepted result. A cheap attempt that needs another pass changes that accounting.
Reasoning changes the integration
Reasoning cannot be disabled. In the Anthropic-compatible interface, sending a disabled thinking setting returns HTTP 400. The effort control accepts levels from low through max. It defaults to maximum reasoning effort. Higher effort generates more reasoning tokens and increases latency.
The same API reference says thinking tokens count toward the maximum generation budget. A budget that's too small can end the response without any answer text. Developers adapting an existing client should therefore check its output limit alongside its reasoning settings, especially if the client assumes that every successful response includes a visible answer.
Those are documented behaviors, not changes we have traced to this release. They explain why the launch's token rates alone can't predict the duration or bill for a complete coding task. Testing low and max effort against the same acceptance checks would show whether additional reasoning earns its cost on that work.
Weights are still a separate milestone
MiniMax's release-note index calls M3.1 Flash an open-weight model, while the launch announcement says the weights will be released later. The announcement doesn't give a date or license terms for that release. Readers planning to host it themselves still need the checkpoint and its terms before they can assess deployment. For API users, the next useful evidence is independent testing that records effort settings and the cost of completed tasks.