MiniMax M3.1 Flash Tops Business Benchmark but Carries a 93% Hallucination Rate

Quick Facts

• MiniMax M3.1 Flash scored 62.2 on Automation-Bench v1.0.6, ahead of DeepSeek V4.1 Flash-1017 (60.5) and Claude Opus 5 (50.3).
• The model costs $0.15 per million input tokens, about 57% cheaper per task than its predecessor M3.
• Independent testing found a 93% hallucination rate because the model attempts nearly every question rather than declining to answer.

MiniMax released M3.1 Flash on October 19, marking the formal launch of a model that had already been running in the wild for nearly a month under the name Space Bunny Alpha.

The stealth version appeared on OpenRouter on September 23 under the provider tag "Stealth." Developers spent two weeks reverse-engineering its origins before MiniMax confirmed authorship. During that anonymous run, Space Bunny led OpenRouter for 10 days with 56.7 trillion tokens processed and topped OpenCode for 12 days with 106 trillion tokens.

The public release is meaningfully different from the stealth version. Feedback from that anonymous run let MiniMax fix bugs mid-deployment. Time-to-first-token dropped from 1.95 seconds to 1.22 seconds, throughput climbed from 70 to 100 tokens per second, and the error rate fell from 1.4% to 0.7%.

Architecture and Speed

M3.1 Flash is a 428-billion-parameter mixture-of-experts model with roughly 23 billion active parameters. It supports a 1-million-token context window and accepts text, image, and video input. MiniMax says the model uses a more aggressive sparse attention architecture, with output speeds exceeding 200 tokens per second in daily use and a median rate above 170 tokens per second.

The company describes thinking as always on, with five selectable effort levels. A Pro version aimed at higher intelligence is planned but has no release date.

Benchmark Results

On MiniMax's own benchmarks, M3.1 Flash scored 62.2 on Automation-Bench v1.0.6, ahead of DeepSeek V4.1 Flash-1017 at 60.5, Claude Opus 5 at 50.3, and GPT-5.6 Sol at 45.8. On BankerToolBench, which tests multi-step financial planning and consecutive tool calls, it scored 76.2, tied with Claude Opus 5.

Claude Opus 5 still leads on Terminal Bench 4.0 with 51.8 versus M3.1 Flash's 38.1, and on SWE Refactor Bench with 31.0 versus 23.0. On DeepSWE v1.1, DeepSeek leads with 77.0 against M3.1 Flash's 73.7.

Artificial Analysis tested M3.1 Flash independently and placed it at 42 on the Artificial Analysis Intelligence Index, up 13 points from M3 and level with GLM-5.3-Flash, just ahead of Gemini 3.8 Flash at 41, and behind DeepSeek V4.1 Flash-1017 at 45. On complex reasoning via Humanity's Last Exam, M3.1 Flash scored 48%, ahead of DeepSeek at 43% and GLM-5.3-Flash at 40%.

The Hallucination Problem

The most significant concern from independent testing is reliability. On the AA-Omniscience factual accuracy test, M3.1 Flash answered 56% of questions correctly, well above DeepSeek at 43% and GLM-5.3-Flash at 28%. But because the model attempts nearly every question instead of declining uncertain ones, Artificial Analysis recorded a hallucination rate of 93% under the benchmark's definition, producing an overall AA-Omniscience score of just 16.

The benchmark counts an attempted-but-wrong answer as a hallucination regardless of raw accuracy. The practical risk for enterprise buyers is clear: a model that never says it does not know will produce confident wrong answers at scale.

Pricing and Cost

M3.1 Flash is priced at $0.15 per million input tokens and $0.60 per million output tokens. Cached input drops to $0.015 per million. That is half the per-token price of M3 and matches DeepSeek's standard rates for V4.1 Flash, though DeepSeek's off-peak cached input rate of $0.003 is lower.

Artificial Analysis calculated the average cost to complete an Intelligence Index task at $0.22 for M3.1 Flash, down from $0.51 for M3 and slightly below DeepSeek's $0.26 per task. On Automation-Bench, M3.1 Flash reaches a 62.8% completion rate at roughly $0.04 per task, compared to nearly $1 per task for Claude Opus 5.5 at similar performance levels.

Company Background

MiniMax was founded in December 2021 by Yan Junjie, a former SenseTime vice president. The Shanghai-based company closed a $300 million Series B extension in July 2025 at a $4 billion valuation, led by Shanghai state-owned capital through Shanghai STVC Group, bringing total funding to approximately $1.15 billion. MiniMax listed on the Hong Kong Stock Exchange in January 2026.

For software teams evaluating the model, M3.1 Flash offers strong agentic performance and low cost per task. Deployments that require a model to acknowledge uncertainty, such as customer-facing tools or compliance workflows, will need additional guardrails given its tendency to answer rather than abstain.