MiniMax M3.1 Flash takes the agentic cost lead
The lead on the agentic cost curve now belongs to MiniMax M3.1 Flash.
| MiniMax M3.1 Flash | |
|---|---|
| API price | $0.15 / $0.60 per 1M input/output tokens |
| cache read | $0.015 / 1M tokens |
| Automation-Bench | 62% |
| ALE | 30% |
| DeepSWE | 74% |
| BabyVision | 91% |
| vs M3 | 1/3 the cost of M3 and 3x the speed |
| weights date | Oct 23 |
The model is listed at $0.15 per 1M input tokens and $0.60 per 1M output tokens, with cache reads at $0.015 per 1M tokens. Weights are set for Oct 23.
The model carries 428 billion total parameters and 23 billion active parameters.
The benchmark lead is the change. MiniMax M3.1 Flash leads on Automation Bench, Banker ToolBench, and LOCA-Bench-256K, while DeepSeek-V4.1-Flash-1017 stays ahead on the DeepSWE cost curve.
The per-token compute reduction is claimed at one-twentieth of the previous version. Distillation does not explain the disparity. Independent verification of the benchmark scores and the training compute is still missing.
The cost lead is the change; the missing baseline is the open question.
After @MiniMax_AI. We did not report this. The pictures, if any, are theirs.
