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

Same prompt for a Three.js koi pond with waterfall and feed button: MiniMax finished in 3.3 minutes for $0, DeepSeek took 1h40m and $0.9; the author found DeepSeek's pond prettier but MiniMax's fish weaker.

Running the same four parallel agent jobs in native apps, M3.1 Flash took 19:52 at roughly $0.30 in estimated token cost versus 10:57 and at least $4.52 for Opus 5.5.

Having used the anonymous Space Bunny last month, the author confirms it was M3.1 Flash: $0.15/$0.60 per million tokens, 3x faster, strong on coding and agentic work, versus 25-30x pricier Opus 5.5 and GPT-6 Astra.

A user has M3.1 Flash generate a first-person arena FPS entirely in code with three.js and no external assets, calling the progress clear while suspecting benchmark gaming.

The author used the anonymized Space Bunny in OpenCode to build a complete browser game, SkyGarden, from a single prompt in minutes, and after the reveal called it impressive for $0.15/$0.60 per million tokens at 3x M3's speed and one-third the cost.

A subscriber reports strong task performance and native multimodal vision, noticeably ahead of GLM-5.3-Flash on the same reverse-engineering task, plus good safety behavior.

A roundup of this week's open-weight releases noting MiniMax M3.1 Flash and DeepSeek V4.1 Flash share $0.15/$0.60 pricing (DeepSeek doubles at peak, MiniMax flat), that M3.1 Flash's early anonymous version was Space Bunny, plus a 50-PR benchmark of Meta Muse Spark Open.

Both Flash models cost $0.15/$0.60 per million tokens; DeepSeek leads DeepSWE 77.0 to 73.7, while MiniMax claims 200+ tok/s, OSWorld V2 61.7, pricier cache and weights still pending.

In a FAWNBENCH head-to-head, DeepSeek V4.1 Flash cost $0.067 and took about 5 minutes while MiniMax M3.1 Flash cost $0.029 and took about 3 minutes, surprising the author.

The author had pitted the anonymous Space Bunny against Opus 5 and only noticed it was unusually fast; they now learn it was an early MiniMax M3.1 Flash, whose official release scores 62.2 on Automation-Bench ahead of Opus 5.

Same task test: MiniMax M3.1 Flash took 90 minutes and $0.50 with a simpler design, while DeepSeek V4.1 Flash took 80 minutes and $1.15; both were judged too slow for Flash models.

Running the same prompt on both Flash models, MiniMax took 1h30m at about $0.50 with a simpler design, while DeepSeek took 80 minutes at $1.15; the author notes both were far too slow for Flash models.