MiniMax released M3.1 Flash on October 19, 2026, with access through its API, M Plan and MiniMax Code. Its launch announcement says model weights will follow later. Hosted access should not be confused with an already completed downloadable release. See the launch announcement and dated release notes.
For teams evaluating agent workflows, our assessment is that the useful question is not simply whether tokens are cheaper. It is whether the model can produce an accepted result at a lower total operating cost.
What the published API rates show
MiniMax 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. Its standard M3 rates for inputs up to 512,000 tokens are $0.30, $1.20 and $0.06 respectively. These are published unit rates, not a promise about your workflow's total bill. Source: MiniMax pricing.
The configuration detail to check first
The invocation guide says thinking defaults to max. Higher effort produces more output tokens and higher latency. Thinking cannot be disabled; reducing effort is the documented alternative. A useful comparison should therefore record the effort setting rather than silently comparing defaults.
How we would evaluate a switch
Start with one bounded workflow whose correct outcome can be checked. Keep the task set, available tools, permissions and acceptance criteria consistent between candidates. Include ordinary cases, ambiguous requests and tasks that should be escalated rather than executed.
Measure total model spend across every attempt, time to an accepted result, tool failures and reviewer effort. Report failed runs alongside successful ones. Otherwise, a cheap successful example can conceal an expensive retry pattern.
Our suggested decision metric is total evaluation spend divided by accepted outcomes, with review time reported separately. Do not treat that as a vendor benchmark. It is a proposed way to make your own comparison more relevant to operations.
What this launch does not settle
Lower published rates do not establish that a migration is worthwhile. Require your own acceptance threshold before expanding access, and keep rollback available. For self-hosting plans, verify an actual downloadable release and its license rather than relying on an open-weight label.
The next step is a narrow experiment with an explicit stop condition, not an immediate replacement of every model in an agent stack.