MiniMax released M3.1 Flash on October 19, 2026, and it immediately posted a 62.2 on Automation-Bench v1.0.6, clearing both DeepSeek V4.1 Flash-1017's 60.5 and Claude Opus 5 on a benchmark designed around real SaaS workflows across sales, marketing, operations, support, finance, and HR. The price tag is the headline: $0.15 per million input tokens, $0.60 per million output tokens.

Artificial Analysis puts a full benchmark run at roughly $0.22 per task, down from about $0.51 on the predecessor M3. That's a 57% cut in evaluation cost per task. MiniMax's own claim is sharper: 70% less cost on agentic workloads versus M3, with inference nearly three times faster and median output above 170 tokens per second.

The 428-billion-parameter model carries a one-million-token context window and lands at an Intelligence Index of 42, up from 29 on M3. For reference, GLM-5.3 Flash also sits at 42 (around $0.28 per task), DeepSeek V4.1 Flash-1017 at 45 (around $0.26), and Gemini 3.8 Flash at 41 but roughly $1.24 per task, the highest-cost option in the cluster by a wide margin.

Then the asterisk.

On AA-Omniscience, M3.1 Flash answered 56% of questions correctly, versus 43% for DeepSeek V4.1 Flash-1017 and 28% for GLM-5.3 Flash. Overall AA-Omniscience score: 16. The structural problem shows up in the error profile. Of the questions it got wrong, 93% were answered confidently rather than refused. For an agent making dozens of tool calls inside a sales or marketing workflow, that's the difference between a model that fails loudly and a model that fails plausibly.

MiniMax's own reported numbers fill out the picture: 73.7 on DeepSWE v1.1 (behind DeepSeek V4.1 Flash-1017's 77), 76.2 on BankerToolBench (tying Opus 5), 61.7 on OSWorld 2.0, 29.6 on Agents' Last Exam, and 82.2 on CyberGym. The model also surfaced earlier as "Space Bunny," an anonymous entry on OpenRouter that processed more than 200 trillion tokens in late September 2026 before attribution.

The frame worth holding onto is the one Gemini 4 Argon's Automation-Bench debut made legible earlier this month: business-workflow benchmarks have become the terrain where frontier labs and budget-tier Chinese models trade positions week to week. M3.1 Flash is the first budget-tier entrant to take the Automation-Bench lead at this price point, and the weights are scheduled to drop on October 23, which will add self-hosting economics and a license to the conversation.

Until then, the honest deployment pattern is supervised. A model that's confidently wrong 93% of the time it's wrong doesn't belong at the end of an unattended pipeline; it belongs upstream of a human verification gate.

MiniMax M3.1 Flash tops Automation-Bench at $0.15 per million input tokens — with a 93% hallucination caveat

Sources