Article 305: MiniMax M3.1 Flash Sells Agent Work at $0.15 per Million Input Tokens - and Was "Space Bunny" All Along

MiniMax released MiniMax-M3.1-Flash on 19 October at $0.15 per million input tokens, $0.60 per million output and $0.015 per million for cache reads, single tier, with China pricing at 1 yuan input, 4 yuan output and 0.1 yuan for cache reads. The model is a "frontier multimodal coding model" with a 1M context window, native text, image and video input, and tunable thinking depth via an effort field accepting low through max, defaulting to max. It always thinks: asking the Anthropic-compatible endpoint to disable thinking returns HTTP 400. The company claims online inference nearly three times faster, output above 200 tokens per second with a median above 170, and 70% lower overall cost on agentic tasks than its predecessor. Reported benchmarks include DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7, Agents' Last Exam at 29.6, BankerToolBench at 76.2 and CyberGym at 82%.

Then the reveal. Space Bunny - 200 trillion tokens processed while running anonymously - was an early version of M3.1 Flash. OpenRouter says it was its number one model for ten straight days and served 56.7 trillion tokens before leaving on 5 October; OpenCode says it was number one there for twelve consecutive days and served 106 trillion. Independent standings are less triumphant: Arena's Code Arena WebDev leaderboard has M3.1 Flash at 1,588 points, 12th among open models and 32nd overall, up 106 points from MiniMax-M3. Weights arrive 23 October per MiniMax's release post; free access is running on OpenRouter, OpenCode and Cline, and Ollama Cloud lists it at launch with US hosting and zero data retention.

Key Takeaway: The anonymous-model trick is now a repeatable go-to-market motion, and its implications are worse than a marketing stunt. A lab can put an unbadged model in front of developers, harvest weeks of unvetted usage in agentic harnesses, and only reveal whose weights those were after the leaderboard position is already earned. It works because agent developers optimise for cost per completed task, not for vendor identity - 56.7 trillion and 106 trillion tokens of traffic were the real benchmark, and they were bought with free access. The 106 points of Arena gain over the previous variant, with a 32nd-place overall finish, is the honest scorecard: excellent at the price, not frontier. That gap is the whole Chinese open-weight strategy in one line - sell competent at a fifth of the cost and let volume do the arguing. Two dates decide how it lands: 23 October, when weights drop and the price becomes negotiable, and 18 November, when M2.5 and M2.7 retire and any pipeline still pinned to the old names stops answering.

Sources: MiniMax research blog, 19 October 2026; DigIntoAI, updated 20 October 2026; SiliconANGLE; Arena Code Arena leaderboard