MiniMax rolled out M3.1 Flash today, October 19, and the pitch is simple: near top-tier agent work for a fraction of the price.
The highlights
- Price: $0.15 per million input tokens, $0.60 output, $0.015 cached input. MiniMax says agent tasks cost about 70% less than its last Flash model.
- Speed: close to three times faster online inference, with output over 200 tokens per second and a median around 170.
- Benchmarks (MiniMax’s numbers): 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, and 82% on CyberGym for finding and validating vulnerabilities.
- Stealth test: an early version ran anonymously as “Space Bunny” and processed more than 200 trillion tokens.
- Availability: MiniMax API, M Plan and MiniMax Code now. Weights come later, and a bigger Pro model is on the way.
My take
The price floor for agent models keeps dropping. These are vendor benchmarks, so I want to see independent numbers, but if those open weights land it could be another good option for local and self-hosted agent runs.
Have a blessed day! And Happy Hacking!
