MiniMax launched its updated M3.1 Flash AI model on 19 October, claiming improvements in agentic tasks, programming and cybersecurity, Reuters reported from a company statement. MiniMax also said online inference had become nearly three times as fast and usage costs had fallen by 70%. The named launch follows an anonymous listing that drew attention from developers trying to identify the model behind it.
Key points
- MiniMax claims gains in agentic tasks, programming and cybersecurity.
- It says online inference is nearly three times as fast, with usage costs down 70%.
- Space Bunny Alpha appeared anonymously on OpenRouter before a M3.1 Flash preview entered MiniMax Code.
- No published benchmark score for M3.1 Flash accompanied the preview.
Space Bunny appeared before the MiniMax name
On 23 September, a free model named Space Bunny Alpha appeared on OpenRouter under a provider label reading “Stealth”, Startup Fortune reported. Its listing offered a 1 million token context window and video-reading capability. Developers compared its outputs and tried to identify its maker while the listing remained available. It expired on 5 October.
Space Bunny was M3.1 Flash, Startup Fortune reported, attributing the identification to comparisons of its behaviour and model fingerprints with MiniMax’s existing family. That account connects the anonymous model with the named release, but the activity on OpenRouter was developers trying the system on their own tasks, rather than a published evaluation of the capabilities MiniMax claimed at launch.
MiniMax had also placed M3.1-Flash-Preview in MiniMax Code on 27 September, available to people already on a Token Plan subscription. Startup Fortune reported that this preview arrived without a press release, model card, benchmark report or public price. The sequence matters for anyone building with it: the September preview gave access to a named model, while the October announcement carried MiniMax’s speed, cost and capability claims about the updated M3.1 Flash model.
MiniMax’s speed and cost claims
MiniMax described the gains in agentic tasks, programming and cybersecurity as “generational”, according to Reuters. An agentic coding task can send a model through repeated rounds of calling tools, changing files and running tests, rather than asking for one answer and stopping there. Each round uses tokens, and each wait for a response adds to the time spent on the task. That makes inference speed and usage cost particularly consequential for this kind of work, provided the responses are useful enough to keep the work moving.
Editing a file could become quicker when a requested change, a test and a correction each require another response. If MiniMax’s claimed speed and cost changes carry through to those repeated exchanges, they would make such work less expensive and less interrupted by waiting. A change that needed further correction would still take another round.
The prices Startup Fortune gives for MiniMax’s full M3 model are $0.30 per million input tokens and $1.20 per million output tokens. Input tokens are the material sent to a model; output tokens are what it generates in reply. Those are prices for full M3, not a published price for the M3.1 Flash preview. For a developer comparing models, the distinction matters because a percentage reduction in usage costs is a different piece of information from a price for the particular model being used.
M3.1 Flash awaits a published benchmark score
MiniMax has cited an 80.5% result for full M3 on SWE-bench Verified in making its case for that model’s coding performance. That is a company-reported score for full M3. No published benchmark score accompanied M3.1 Flash at its preview release, and no widely published, weighted score for M3.1-Flash-Preview had appeared as of 19 October.
One narrower result came from a private ModelDial evaluation described by Startup Fortune. On a backend-and-testing coding task, M3.1-Flash-Preview received 30 out of 100 when run at low reasoning effort. Reasoning effort is a setting that lets a developer vary how much computation the model uses for a task. The reported score belongs to that task and setting; MiniMax’s launch claims concern a broader set of agentic, programming and cybersecurity uses.
