Image: MiniMax
MiniMax has launched M3.1 Flash, a cheaper, faster version of its M3 model built for AI agents that handle office work and write code. It also cleared up a small mystery: Space Bunny, an anonymous model that MiniMax says processed more than 200 trillion tokens while its maker stayed hidden, was an early version of M3.1 Flash.
MiniMax M3.1 Flash: Space Performance + Bunny Speed
— MiniMax (official) (@MiniMax_AI) October 19, 2026
🚀🰠Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence… pic.twitter.com/j0tCfNYYgo
The pitch is all about cost. MiniMax charges a flat $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015. It says agent tasks cost 70% less than on M3 and run nearly three times faster, with typical output speeds above 170 tokens per second.
Part of that saving is simply a lower sticker price. OpenRouter lists M3 at $0.30 and $1.20, exactly double. If MiniMax's figures hold, the rest has to come from the new model using fewer tokens to finish the same job.
Here's the detail MiniMax doesn't put up front: those are exactly the prices DeepSeek charges for its V4.1 Flash 1017 model outside peak hours. DeepSeek doubles its rates during two weekday windows that line up with Beijing working hours (6:30 to 9:30 AM and 11:30 AM to 3:30 PM IST), and MiniMax's own cost chart prices DeepSeek at those peak rates. So the "flat rate" is mainly an advantage if you work during those hours.
The benchmarks tell a similar story. In MiniMax's own comparison table, M3.1 Flash wins the office-automation tests, scoring 62.2 on Automation-Bench against 60.5 for DeepSeek V4.1 Flash 1017 and 50.3 for Anthropic's Claude Opus 5. But DeepSeek's model beats it on several coding tests, including DeepSWE (77.0 vs 73.7), Terminal Bench 4.0 (42.3 vs 38.1) and NL2Repo (68.2 vs 61.2). Opus 5 still leads most rows outright.
Image: MiniMax
MiniMax's DeepSWE chart is just as candid. It shows DeepSeek's model reaching higher scores at a lower cost per task. Against pricier rivals it shines: its top setting roughly matches Claude Opus 5's best DeepSWE score for a tiny fraction of the cost per task.
Image: MiniMax
All these numbers are MiniMax's own. Artificial Analysis, an independent tester, hasn't scored M3.1 Flash yet. It rates M3 at 29 on its intelligence index and DeepSeek V4.1 Flash 1017 at 45, so Flash has a big gap to close if MiniMax's claims are to hold up. Space Bunny's 200 trillion token figure can't be checked either.
MiniMax's X post promises open weights on October 23, though its blog only says "later." MiniMax hasn't said how big the model is. M3 has about 428 billion parameters, with 23 billion active at a time, which puts it in server territory rather than on a home PC.
The licence matters too. M3 shipped under the MiniMax Community License, which requires "Built with MiniMax M3" credit for commercial use and written permission for businesses making more than $20 million a year. DeepSeek's model uses the far looser MIT licence. MiniMax hasn't said which terms Flash will get.
Price fights matter for MiniMax's business. The company, listed in Hong Kong since January, nearly quadrupled its revenue to $116.6 million in the first half of 2026, and 63.4% of that came from its API and enterprise services. It's still losing money, with a $293 million adjusted net loss. MiniMax says a more powerful Pro model is coming later.
How to try it
M3.1 Flash is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. OpenRouter is also listing a free version for now. Since it costs the same as DeepSeek's model off-peak, it's worth running a few of your own real tasks on both before switching.
Comments