MiniMax M3.1 Flash scores 62.8% on Automation-Bench, ahead of Opus 5.5 at far lower cost per task
Hands-on post reports M3.1 Flash at 62.8% on Automation-Bench (ahead of Opus 5.5), 76.2 on BankerToolBench and 61.7 on OSWorld V2, and shares a prompt that runs an agent three times and shows the cost of each run.
MiniMax M3.1 Flash is bad news for every expensive model It’s the model that was secretly running as Space Bunny, and its agent results explain why so many people were already building with it 62.8% on automation-Bench, ahead of Opus 5.5 at a fraction of the cost per task, 76.2 on BankerToolBench for long multi-step tool chains and 61.7 on OSWorld V2 for real computer use That matches what I’ve seen building agents. The model is rarely what breaks first, the bill is. Long tool chains, retries and self-checks add up fast, and I used to cut steps just to keep a run affordable When every run costs this little you stop rationing and let the agent do the job properly: plan, retry, verify, run it again Send this prompt to M3.1 Flash (no usage cap on MiniMax Code right now) and it automates your most boring weekly task and shows what every run cost: “Take the task I repeat every week. Build an agent that does it end to end, run it three times, and show me the cost of each run with the math.” Now you know what your agent really costs, and you saw it yourself instead of trusting a pricing page... The full prompt and the chart are below, try it on your most annoying task tonight
This post quotes another post: x.com/minimax_ai/status/2112150867868541221