MiniMax rolled out M3.1 Flash on October 19 and used the launch to clear up a mystery. Space Bunny, an anonymous model that held OpenRouter’s top spot for ten straight days, was an early build of the same system.
OpenRouter said Space Bunny processed 56.7 trillion tokens during the stealth run, the most of any anonymous model it has hosted. MiniMax’s own figure for the build was over 200 trillion tokens across its evaluation work.
The model is a 428B-parameter mixture of experts with roughly 23B parameters active and a 1M-token context. Pricing is flat at $0.15 per 1M input tokens and $0.60 per 1M output, with cache reads at $0.015. There is no long-context surcharge.
Benchmarks land where vendors usually put them
MiniMax reports 62.2 on Automation-Bench v1.0.6 and 76.2 on BankerToolBench, tying Claude Opus 5 on the second. DeepSWE v1.1 comes in at 73.7, above GPT-5.6 Sol at 73.0.
Coding is the weak spot. Terminal-Bench 4.0 sits at 38.1, far behind Opus 5, and SWE Refactor Bench reaches 23.0 against 31.0 for Opus. Every figure is vendor-reported and none has independent replication.
Reasoning is always on and cannot be disabled, with five effort levels from Low to Max. Requests that try to switch thinking off return an error, so teams migrating from M3 need to update parsers that read inline thinking tags.
OpenRouter is running a 72-hour free window. Weights are due October 23 with the license undisclosed, which keeps self-hosting plans on hold.
