MiniMax M3.1 Flash Revealed: the Anonymous 'Space Bunny' Goes Public at 200 TPS

MiniMax released M3.1 Flash on October 19, and the release note contains a disclosure strategy more labs should study: the model has been serving traffic for weeks under the anonymous alias “Space Bunny,” during which it processed over 200 trillion tokens of real usage before MiniMax put a name on it. The pitch is agentic work at commodity prices: 200+ tokens per second in everyday use (170+ at the median), nearly 3x faster online inference than its predecessor, flat pricing of $0.15 per million input tokens and $0.60 per million output tokens with cached input at $0.015, and a claim that agentic-task usage costs 70% less than the prior model. The API, M Plan, and MiniMax Code have it today; weights are promised for later release, with a more powerful Pro variant teased.

What the benchmarks say, with the caveats attached

The card leads with agentic evaluations rather than chat scores: BankerToolBench at 76.2, DeepSWE v1.1 at 73.7%, and first-place-among-Chinese-models results on OSWorld V2 (61.7) and Agents’ Last Exam (29.6). CyberGym sits at 82%, and MiniMax says internal expert reviews single out code auditing, penetration testing, and reverse engineering. Two honest caveats belong next to those numbers: MiniMax disclosed no parameter count or architecture details, mentioning only “more aggressive sparsity” for inference efficiency, so there is no way yet to situate M3.1 Flash against the open-weights field on a per-compute basis; and “first among Chinese models” is a carefully chosen frame, with DeepSeek’s V4.1-Flash-1017 refresh having posted 42.3 on Terminal-Bench 4.0 a day earlier under different harness conditions. The Space Bunny history is the most credible benchmark of all: 200 trillion tokens of live traffic is a harder test than any leaderboard.

The pricing ladder in Chinese AI is collapsing

Set the pricing against the month’s moves and a pattern emerges. DeepSeek’s Search and Fetch APIs repriced web tools at $0.75 per 1,000 successful calls; MiniMax now sells agentic model tokens at $0.15/$0.60 with a cached-input tier at 10% of list; and both firms are racing toward cost-per-completed-task as the headline metric rather than cost-per-token. The “70% lower agentic-task cost” claim is the tell: agents burn tokens on retries, tool formatting, and context re-reading, so a model that is merely cheap per token can still be expensive per finished task. MiniMax is claiming the whole curve moved. For engineering teams running multi-step workflows, the practical consequence is that model cost is approaching the floor where orchestration overhead, latency budgets, and failure handling dominate the bill; the month’s efficiency research points the same direction.

The anonymous-release playbook, and what to watch

Shipping under an alias before the reveal is becoming a recognizable move: it buys weeks of production hardening and generates an unpolluted usage record, then converts that history into launch credibility. The risks are symmetric: users who picked Space Bunny for the wrong reasons (or avoided it on identity grounds) got no notice, and enterprise buyers generally require disclosure of who is processing their data. What to watch: the weights release, which will determine whether M3.1 Flash becomes a self-hosting option or stays an API-only product, since the license terms are not yet stated; the promised Pro variant, which will show whether MiniMax is competing at the frontier or ceding it; and independent reproduction of the CyberGym security numbers, because a model marketed partly on offensive-security skill is a claim the community should verify with more than the vendor’s expert reviews.