MiniMax launches M3.1 Flash model, says it cuts AI agent costs by 70%

MiniMax has released M3.1 Flash, a faster and cheaper follow-up to the large language model it launched in June. The Chinese developer says the new version is aimed squarely at AI agents, the sort of systems that burn through dozens of model calls before they finish a single task.

The company says agentic work now costs 70% less than on its previous model, while inference runs close to three times faster. At peak everyday usage, it says output can top 200 tokens per second. The model’s size has barely moved from M3: MiniMax’s preview guide lists 428 billion parameters, and the extra speed comes from more aggressive sparsity at inference time.

This is still a reasoning model on every request, but developers can choose from five effort levels, from Low to Max. M3.1 Flash accepts text, image and video input, and prompts can stretch to 1 million tokens. MiniMax also put up benchmark numbers: 61.7 on OSWorld V2, 76.2 on BankerToolBench, and 82% on CyberGym. The company says no Chinese model has done better than its OSWorld score.

Price is part of the point. MiniMax is charging 15 cents per million input tokens, which puts M3.1 Flash directly against DeepSeek’s V4.1-Flash at the same off-peak rate. On DeepSWE v1.1, MiniMax says its model scored 73.7%, while DeepSeek said its own update resolved 77% of tasks over the weekend. Neither result has been independently checked.

The launch also ties off a small mystery. Space Bunny, the anonymous model that appeared on OpenRouter in late September, was actually an early M3.1 Flash build, MiniMax said. More than 200 trillion tokens passed through it while it was still unnamed. M3.1 Flash is available now through the MiniMax API, its M Plan subscription and the MiniMax Code desktop app, with output priced at 60 cents per million tokens and cached input at 1.5 cents. Flash weights are coming later, and a more capable Pro model is on the way.