MiniMax introduced M3.1 Flash on October 19, 2026, with access through MiniMax API, M Plan and MiniMax Code. Announced per-million-token rates are $0.15 input, $0.60 output and $0.015 cached input. Claimed 70% agent-cost savings against its predecessor are vendor results, not SIExpo tests.

Access and weights are separate questions

The launch announcement says weights will be released later. Meanwhile, the model-invocation guide calls it open-weight. Given that inconsistency, we do not treat the descriptive label as proof that downloadable files for self-hosting are already available. We give no unverified date for their publication.

Our conclusion for a team choosing a supplier is to separate two decisions: trying a hosted service and planning a local installation. One can be assessed through an operational interface and its terms; the other requires actual files, a license and technical requirements. A product description does not fill those gaps. A deployment plan should identify which of those access paths it is depending on.

The API guide describes a one-million-token context and five effort levels, with max as the default. It shows Anthropic-compatible and OpenAI-compatible interfaces. That allows planning an integration trial, but compatibility of the request format does not establish identical behavior for every field or tool.

The cost of a completed task

Our example is an agent preparing a change in a small software project. Its completion criterion would be a specific working fix that passes preselected checks. An inexpensive generated answer without a working change does not meet that criterion. This is a proposed evaluation method, rather than a result we obtained with MiniMax.

Such a trial should record the total calls, all billed tokens and time spent on human completion. Equivalent tasks could then be compared under the same permissions and budget. If the effort setting changes, that choice should remain visible in the result. Otherwise, two runs would no longer describe the same operating conditions, even if they share a model name.

A long context should also be evaluated against a concrete need: whether the agent locates the right file and preserves important constraints in the proposed fix. Input capacity by itself does not measure successful work across an entire project. A team might include both a concise case and a case requiring material from several files, while keeping the required outcome explicit.

Retaining unsuccessful attempts would make the comparison more useful. They can show whether spending was lost to repeated corrections, a wrong initial assumption or an unresolved requirement. Those categories are proposed diagnostic records, not an assertion that this particular model suffers from them.

M3.1 Flash adds a hosted option that teams can evaluate. Practical value depends on spending per reproducibly correct outcome. Self-hosting remains a separate question requiring verification of actually published weights and their license, rather than an inference from an inconsistent label.

Editorial verification: 2026-10-20T06:10:23.483994+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.