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

On the same Blender Attack on Titan Shiganshina prompt, the author found M3.1 Flash better on building, cloud and wall detail and overall look, with GLM 5.3 Flash also solid; both cost under $5.

Hands-on test in MiniMax Code: one prompt produced a single-file HTML Japanese train station diorama in 17 minutes, 66k output tokens, $0.04, no fixes needed, with a full train cycle, crossing arms, drifting sakura petals and day/night toggle.

The author links MiniMax's falling gross margin to M3.1 Flash's nearly 3x faster inference, arguing efficiency must pay for price cuts and H2 margin is the number to watch.

The same prompt for a Mona Lisa with a cross and SLAM is run through Claude Opus 5.5 and MiniMax M3.1 Flash, with the two outputs shown side by side.

An anonymous model called Space Bunny led OpenRouter for 10 days and burned 56.7 trillion tokens before being revealed as MiniMax M3.1 Flash at $0.15 per million input tokens; the author then gave it a real job.

The author says three prompts and two hours of MiniMax M3.1 Flash produced a browser open world with cops, helicopters and a heist, at 200+ tokens/s and $0.15/M input, $0.60/M output, for $0.60 total.

The author tested M3.1 Flash on an ambitious coding task, building an interactive 3D city inside a colossal whale skeleton in the Mariana Trench, combining cinematic visuals, camera controls and lighting modes in a single HTML file.

On OpenRouter, M3.1 Flash accepts a 1M-token multimodal context in one call at $0.15 per 1M input and $0.60 per 1M output, with MiniMax reporting gains over M3 on agentic, coding and security benchmarks.

Recreating a koi pond scene with M3.1 Flash, the author found it catches and corrects its own mistakes better and recovers from drift, at $0.15/$0.60 per million tokens and roughly 70% lower cost on agentic tasks than M3.

In a mini ice-rink generation comparison, M3.1 Flash cost $0.21 versus $7.88 for GPT-6 Astra — about 38x — and the original author preferred the cheaper result.

The same prompt produced a pixel skate park in 17 minutes for $0.40 with M3.1 Flash versus 42 minutes and $1.40 with DeepSeek V4.1 Flash, whose darker scene hid half the skaters.

At max effort the model autonomously scraped services and phones from 31 dentist sites, flagged those without online booking, wrote openers, self-checked bugs and verified its report in a browser, all for $0.40 across 20M tokens.