MiniMax launches M3.1 Flash model, claims 70% lower agent task costs

MiniMax released M3.1 Flash on October 19, describing it as an update to the M3 model it launched in June. The company says the model targets the cost of running agents, which often make dozens of model calls to finish one job. MiniMax lists 428 billion parameters, reasoning that stays on for every request with five effort levels from Low to Max, and text, image and video input with prompts up to 1 million tokens.

In MiniMax's own testing, M3.1 Flash scored 61.7 on OSWorld V2 and 76.2 on BankerToolBench, and 82% on CyberGym. The company priced input at 15 cents per million tokens, the same rate DeepSeek charges for its V4.1-Flash model off-peak. DeepSeek reported 77% on the DeepSWE v1.1 test, while M3.1 Flash scored 73.7%; neither result has been independently checked.

The model is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Output costs 60 cents per million tokens and cached input 1.5 cents. MiniMax said it will release the Flash weights later and that a larger Pro model is coming.

If you run agents with many model calls per task, M3.1 Flash offers a lower-cost option to benchmark against your current model, though the cost and speed figures come from MiniMax's own testing.

The pricing is the same as DeepSeek's V4.1-Flash for input tokens, so a comparison should weigh output pricing, speed and your own task results rather than list price alone.