MiniMax M3.1 Flash in practice: tests and comparisons with DeepSeek V4.1 Flash

The author runs the same Lunar Lander prompt and says M3.1 Flash is much better than the anonymous Space Bunny version, crediting weeks of continued RL training, with video.

Hands-on look at M3.1 Flash's 3D frontend skill: a single prompt produced a browser scene with a watermill, stone bridge, stream and a 5-stop camera tour, including smoke, window light and pond reflections, for $1.62 total.

The post recaps the launch: Space Bunny was an early M3.1 Flash build, with nearly 3x speed, 70% lower cost, $0.15/M input, a 13-point jump to 42 on the intelligence index, and first place on OSWorld V2's Chinese leaderboard.

A developer whose main model is DeepSeek V4 is considering M3.1 Flash to cut costs, noting monthly AI spending rose from about $10 to $100, and asks whether quality holds up.

On the Zhizhi leaderboard's private question set, M3.1 Flash scored a median 51.76 versus 23.86 for M3 and 56.03 for DeepSeek V4.1 Flash 1017, but nearly aced old questions and nearly failed new ones, raising benchmaxxing suspicions.

The author ran two hard prompts against anonymous Space Bunny in OpenCode: it built a Connect Four engine with alpha-beta pruning and a monitoring dashboard, running its own tests and fixing bugs, with the official release stronger on agentic work.

A roundup of M3.1 Flash specs and AA data: $0.15 per million input tokens, DeepSWE 73.7%, intelligence index 42, about $0.22 per task, plus a 93% hallucination rate caveat.

A hands-on test used Claude Code with MiniMax M3.1 Flash to generate a whale breach 3D scene for about $0.25, reportedly ~8x cheaper than Astra/Sol and better than GPT 6.1 Sol.

On the same task, DeepSeek V4.1 Flash cost $1.05 over 2h19m while MiniMax M3.1 Flash cost $0.79 over 2h40m, prompting the question of which is worth the price.

On the same prompt and task, MiniMax M3.1 Flash took 1h06m51s with 8.12M tokens for $0.22, while DeepSeek V4.1 Flash 1017 took 1h28m with 89.5M tokens for $0.54.

Citing Artificial Analysis, the post notes M3.1 Flash rose from 29 to 42 on the intelligence index and cut cost per task from $0.51 to $0.22, while a 93% hallucination rate undercuts reliability.

With one screenshot and one prompt, stealth Space Bunny built a single-file website from an X profile in 6 minutes using colors from the avatar; the post also cites M3.1 Flash's 73.7% DeepSWE, ~3x speedup and image/video input.