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

API pricing is $0.15/M input, $0.60/M output and $0.015/M cached read, with a 73.7% DeepSWE v1.1 score that edges GPT-5.6 Sol's 73.0%.

The author notes M3.1 Flash is free on OpenCode and OpenRouter yet draws almost no discussion, and points out the mysterious anonymous Space Bunny on OpenRouter was its early version.

Using M3.1 Flash's $0.15/$0.60 pricing with a 9:1 input-output ratio and 90% cache rate, the author estimates Space Bunny's anonymous testing consumed roughly $5M worth of tokens, noting price differs from cost.

The author ran the same prompt at max reasoning on both models: MiniMax M3.1 Flash took 40 minutes at $0, DeepSeek V4.1 Flash took 50 minutes at $0.4; styles differ and old M3 couldn't match this output.

Official specs: DeepSWE 74%, Cowork 62%, BabyVision 91%, $0.15/$0.60 per 1M tokens, one-third the cost and 3x faster than M3, weights open on Oct 23.

MiniMax's Leanna says M3.1-Flash is positioned at the sweet spot between intelligence, speed and cost, and teases that M3-Pro is coming soon.

On one shared test suite, MiniMax M3.1 Flash scores 42 at $0.22 per full run (free on OpenRouter now), versus DeepSeek V4.1 Flash's 45 at $0.26 and GLM-5.3 Flash's 42 at $0.28.

Hands-on: the author notes M3.1 Flash costs half the new DeepSeek Flash and is running the same tests to see if it sticks to facts in product copy, where DeepSeek failed.

The author says the newly launched MiniMax M3.1 Flash is nowhere near deepseek-v4.1-flash-1017 and calls DeepSeek's Flash the best open-source Flash model.

M3.1 Flash is the anonymous Space Bunny (12 days atop OpenCode): 1M context with text/image/video input, one-third M3's cost and 3x speed; 62.2% Automation-Bench beats Opus 5, 73.7% DeepSWE trails DeepSeek V4.1 Flash, ALE ~$111, open weights TBD.

A direct API call with no extra turns or tool calls produced a 3D pagoda scene at max thinking, using about 50K tokens, judged better than DeepSeek V4.1 Flash.

Recaps official specs ($0.15/$0.60 per million tokens, 200+ tok/s, 1M context, open weights Oct 23) and benchmark table, with the author's own Kilnframe run comparing to M3; finds the Flash model beats Opus 5 on some rows but trails on heavy coding, terminal and computer use.