An anonymous model nicknamed Space Bunny handled more than 200 trillion tokens of traffic before anyone knew who built it. MiniMax, the Chinese AI lab, has now said on its own blog that Space Bunny was an early build of M3.1 Flash, which is on sale today through its API, its M Plan subscription, and its MiniMax Code desktop app.

The post carries no date; the release is dated 19 October 2026. The 200 trillion figure is MiniMax’s own count, but it points to sustained production load rather than a lab demo, which is the harder test for a cheap model.

The price is the headline for builders. MiniMax charges $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million for cached input. The post does not state a context length, so teams with long documents must check that before planning around it.

Two claims arrive without a baseline. MiniMax says online inference is nearly three times faster, but never says faster than what. It also says overall usage cost for agentic tasks is 70% lower than the predecessor, without naming that predecessor or showing the task mix. On raw speed, the company reports output of 200 or more tokens per second, with a median of 170 or more, and credits more aggressive sparsity, meaning less of the model does work on each token.

Capability numbers all come from the vendor, and the post cites no independent evaluation. MiniMax reports 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, and 82% on CyberGym. For DeepSWE it says the result sits “alongside top models,” a phrase that gives no ranking. On AutomationBench-AA it says only that the model is among the top performers, with no score attached.

The one explicit ranking needs care. MiniMax says M3.1 Flash is first among Chinese models on OSWorld V2 and Agents’ Last Exam. That is MiniMax’s own comparison, run against a field it narrowed to Chinese models, so it says nothing about where the model stands against American rivals on those tests. The cybersecurity claims rest partly on evaluations by MiniMax’s internal security staff.

The weights are the open question. The post says they “will be released later” and gives no date, and it says a higher-end Pro model will also arrive later, again undated. Until the weights ship, M3.1 Flash is a hosted service: MiniMax sets the price, the uptime, and where your data goes, and you cannot run it on your own hardware.

The gap matters because open weights are what let a startup lock in a model, fine-tune it, and survive a pricing change. A rented API gives none of that, however low the sticker price.

The sensible move this month is a narrow one. Run your own agent workload through the API, measure cost per finished task rather than per token, and compare that against the 70% claim before moving any production traffic. Keep your current model wired in as a fallback until MiniMax puts a date on the weights.

MiniMax, “MiniMax M3.1 Flash: Redefining Agent Performance and Cost,” MiniMax blog, 19 October 2026.