MiniMax put a name to the season's biggest benchmark mystery on Monday. The Shanghai-based lab officially launched M3.1 Flash and confirmed that "Space Bunny" — the anonymous model that quietly processed more than 200 trillion tokens while climbing the open leaderboard — was an early version of it. MiniMax did not explain how the 200-trillion-token figure was tallied.
The new model is a sparse mixture-of-experts system with roughly 428 billion total parameters and about 23 billion active per forward pass, a 1-million-token context window, and native support for text, image and video input. Reasoning is always on and cannot be disabled; developers can only tune it across five effort levels, with Max as the recommended default.
Pricing is the headline for cost-sensitive agent builders: $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 for cached input, flat across the full 1M context with no long-prompt surcharge. MiniMax says overall usage cost is 70% lower than its predecessor while online inference runs nearly three times faster, with output speeds above 200 tokens per second and a median around 170.
The benchmark table is entirely MiniMax-reported and has not been independently replicated. On Automation-Bench v1.0.6, which tests end-to-end enterprise workflows across SaaS systems, M3.1 Flash scores 62.2 — the top result in the company's own comparison, ahead of Claude Opus 5 at 50.3 and GPT-5.6 Sol at 45.8. (The company's cost chart plots the same number as 62.8, a small inconsistency in its own materials.) On the software-engineering benchmark DeepSWE v1.1 it scores 73.7%, edging past GPT-5.6 Sol's 73.0 but clearly behind DeepSeek's V4.1 Flash-1017 at 77.0. Its weakest showing is Terminal Bench 4.0 at 38.1, against 51.8 for Opus 5.
On agentic and computer-use tasks, MiniMax reports 61.7 on OSWorld V2 — the highest among Chinese models in its table, though behind the 65.7 of a frontier US model — and 29.6 on Agents' Last Exam, where it estimates a full 152-task run costs about $111 versus more than $1,000 for the rivals it compared against. It ties Opus 5 on BankerToolBench at 76.2 and scores 82% on CyberGym.
There is fine print worth reading. The comparison chart prices DeepSeek's model at peak-hour rates; DeepSeek's off-peak pricing is identical to MiniMax's, which narrows the cost gap the chart implies. And no third party has yet published a full benchmark run of the final model — the strongest independent signal so far is that Space Bunny's anonymous scores were broadly real.
M3.1 Flash is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Model weights are due October 23 according to the company's X account, though the blog post says only "later" and no license has been announced — the previous M3 shipped under a custom community license rather than an open one, while DeepSeek's competing Flash weights are MIT. A larger "Pro" model "will arrive later" with no date attached.
The launch shows where model competition has moved: the Flash tier is now where agentic economics get decided, and Chinese labs are currently setting the price-performance benchmark that Western labs have to answer. With weights landing Thursday, the next data point will be whether anyone can replicate the numbers.
Comments (0)
Log in to join the discussion
Log InNo comments yet