China's MiniMax promises AI agents 70% cheaper than its previous model

In late September, an AI model with no declared author appeared on OpenRouter, a service that routes developers' requests to different models. Its name: "Space Bunny." Testers traced it back by analyzing how it splits text into tokens. On October 19, the Chinese company MiniMax solved the mystery: Space Bunny was a preliminary version of its new model, M3.1 Flash. According to MiniMax, more than 200 trillion tokens (the units of text a model processes) went through it while it remained anonymous.

M3.1 Flash has been available since October 19 through MiniMax's API, its M Plan subscription and its MiniMax Code programming app. It succeeds M3, launched in June. The vendor promises agent tasks that are 70% cheaper to run than with its previous model, and generation nearly three times faster, at more than 200 tokens per second.

A price matched to DeepSeek's

Pricing is set at $0.15 per million input tokens and $0.60 per million output tokens, and at $0.015 for inputs already in cache. SiliconANGLE notes that this is the input price DeepSeek charges off-peak for V4.1-Flash, the open model DeepSeek unveiled an update to the previous weekend. M3.1 Flash accepts text, images and video, with requests of up to one million tokens. It has 428 billion parameters, according to the developer guide cited by SiliconANGLE. Reasoning is always on, with five effort levels to choose from.

Scores published by the vendor itself

MiniMax reports 61.7 on OSWorld V2, a computer-use test, a score it says no Chinese model has beaten. It also claims 73.7% on DeepSWE v1.1, a programming test, and 82% on CyberGym, a benchmark from the University of California, Berkeley that measures an agent's ability to reproduce real software vulnerabilities. The vendor also says its in-house experts judge the model to perform well in penetration testing and malware analysis, without giving figures.

MiniMax's page does not say who measured these scores, and no independent evaluation of M3.1 Flash was available on October 20. It names no competitor and says only that the model rises to the level of the best. For its part, DeepSeek claims 77% on DeepSWE for V4.1-Flash, also without independent verification.

What to know before trying it

For a European company, two points matter. First, using the API amounts to entrusting your data to a Chinese provider. The documents reviewed do not say where the data is processed: check the terms of use before any professional use. The Space Bunny episode also shows that for several weeks, developers sent their requests to a model whose vendor they did not know.

Second, the model's weights have not yet been published. They would make it possible to run M3.1 Flash on your own servers. MiniMax promises to release them "later," with no date or license. The company, listed in Hong Kong since January 2026, also announces a more powerful Pro version. It remains to be seen whether its scores hold up against the first independent tests.