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

Using OpenCode's free routes at $0 in model fees, both models built their own helicopter with Python-synthesized rotor noise; DeepSeek's first cockpit put glass over a solid wall, and both needed revisions.

Asked M3.1 Flash to generate a Frieren SVG from a reference with self-chosen background music; in Claude Code ultracode mode it took 1h50m and 70M tokens, free via MiniMax plans, and beat expectations.

Same-prompt comparison: M3.1 Flash finished in 48 minutes during its free opencode window, while GLM 5.3 Flash took 1 hour 6 minutes and $3.20; the author found M3.1 Flash more interactive, cleaner in UI and 18 minutes faster.

Using max reasoning and 1M context, the author one-shotted a Flappy Bird clone with parallax scenery, bird tilt, crash screen shake and feathers, live score and replay, taking 38 minutes.

On the same prompt and task, MiniMax M3.1 Flash took 1h 3m 51s with 239.2k tokens for $0.50, while DeepSeek V4.1 Flash 1017 took 1h 12m 43s with 407.3k tokens for $0.90.

A roundup of official specs (428B total / 23B active, 1M context, $0.15/$0.60) and Artificial Analysis measurements: 42 index, level with GLM-5.3-Flash but behind DeepSeek V4.1 Flash 1017, about $0.22 per task, high hallucination rate, weights due October 23.

A user reverse-proxied MiniMax M3.1 Flash out of OpenCode into DSH and says results are much better than using it directly in OpenCode, noting harness choice matters a lot.

Artificial Analysis data puts M3.1 Flash at 42 on the intelligence index, just 3 points behind DeepSeek 1017's 45, while costing $0.22 per task, roughly half as much.

In a Sonnet-generated, web-grounded trivia quiz across four rounds, DeepSeek 1017 scored near 100% while M3.1 Flash scored roughly 60%, 48% and 62%, with a final round at 91% vs 62%.

Before wiring it into an agent: omitting reasoning_effort defaults to max, thinking can't be turned off (disabled/none returns HTTP 400), temperature and top_p are ignored, and the benchmarks are MiniMax's own.

A price comparison of 11 models puts MiniMax M3.1 Flash level with DeepSeek V4.1 Flash 1017 at $0.15/$0.60 per million tokens, with GPT-6 Luna cheapest and Opus 5.5 priciest.

A brief complaint that MiniMax M3.1 Flash is super slow and that GPT-6 Luna feels better, without specific tasks or numbers.