Same prompt test: M3.1 Flash ran 50 commands and delivered nothing, Opus 5.5 built the jelly toy
On an identical sliceable-jelly-toy prompt, M3.1 Flash ran ~50 commands, rewrote the physics engine unprompted, then failed on insufficient balance with zero output, while Opus 5.5 delivered a working toy.
first impression of MiniMax M3.1 Flash vs Claude Opus 5.5 on paper it's not even a fight. MiniMax is 26.7x cheaper on input and 33.3x cheaper on output. so i gave both the exact same prompt: build a sliceable jelly toy. here's what MiniMax did with it > went full agent mode > thought for 10+ minutes > ran ~50 commands > decided "cannon-es is buggy" and started rewriting the physics engine. nobody asked > then died on error 50110: insufficient balance > jelly delivered: zero meanwhile Claude just made the jelly. the pink one in the video. you can actually slice it. wait. their own numbers look great: 73.7% on DeepSWE, 91% on BabyVision, 170+ tok/s, open weights on Oct 23. all self reported, nobody has verified them yet. honest part: this is one prompt, not a benchmark. for small scoped jobs as a subagent i'd still try it, at that price experimenting is basically free. but on anything open ended, 170 tok/s means nothing when it spends them on stuff you didn't ask for. the cheapest model in the room ran out of money first.