MiniMax-M3.1-Flash review: Intelligence Index 42, low cost, but 93% hallucination rate

MiniMax's new Flash reasoning model, MiniMax-M3.1-Flash, received a full evaluation from Artificial Analysis and was added to the Code Arena WebDev leaderboard on October 19. The overall results are strong: an Intelligence Index of 42, 13 points higher than M3, on par with GLM-5.3-Flash and slightly above Gemini 3.8 Flash (41), ranking 4th among 173 comparable models — well above the median of 13. Pricing is $0.15 per million input tokens, with per-task cost around $0.22. On Code Arena WebDev it scored 1588, placing 12th among open-source models and 32nd overall — up 106 points from the previous generation — and lifted MiniMax to 12th among all vendors. The current takeaway: the model delivers first-tier general intelligence at extremely low cost, but reliability is a clear weakness.

Confirmed

• Artificial Analysis Intelligence Index of 42, 13 points above M3, on par with GLM-5.3-Flash, slightly above Gemini 3.8 Flash (41)
• Ranked 4th among 173 models (median just 13); input pricing $0.15 per million tokens, per-task cost around $0.22
• AA-Omniscience knowledge reliability evaluation: 56% accuracy, ahead of GLM-5.3-Flash (28%) and the DeepSeek V4 series, but a hallucination rate as high as 93%
• Mixed agentic results: 73% on AutomationBench-AA, beating GLM-5.3-Flash (max, 60%) and others, but still behind on terminal and real-world tasks
• Artificial Analysis has published full details across all 10 Intelligence Index evaluations (AA-Briefcase, GDPval-AA, Terminal, etc.)
• Code Arena WebDev: 12th open-source, 32nd overall, at 1588 — up 106 points from the previous generation
• User tests show decoding speeds of 68.8–105.7 tok/s, averaging nearly 100 tok/s with no provider errors; rated slightly slower than DeepSeek-v4.1-flash

Why it matters

• Ultra-low pricing plus a first-tier Intelligence Index makes it a strong candidate for cost-sensitive use cases; the drop in per-task cost matters a lot for large-scale deployment
• The split between a 93% hallucination rate and leading agentic benchmarks means users should evaluate by task type: strong on automation tasks, but risky where knowledge reliability is critical

2026-10-19 ~ 2026-10-19 · 9 related posts

• MiniMax-M3.1-Flash hits ~100 tok/s in speed tests, trailing DeepSeek v4.1-flash — TheZachMueller · 2026-10-19
• [source] Code Arena WebDev: MiniMax-M3.1-Flash ranks #12 among open models, DeepSeek-V4.1-Flash tops at 1664 — arena · 2026-10-19
• MiniMax-M3.1-Flash debuts at #12 among open models in Code Arena with 1588 pts — cedric_chee · 2026-10-19
• MiniMax-M3.1-Flash scores 42 on AA Intelligence Index, halving cost per task to $0.22 — ArtificialAnlys · 2026-10-19
• [source] MiniMax-M3.1-Flash leads AutomationBench at 73% but trails on Terminal-Bench — ArtificialAnlys · 2026-10-19
• MiniMax-M3.1-Flash tops knowledge accuracy but hallucinates on 93% of unknowns — ArtificialAnlys · 2026-10-19
• Full per-evaluation breakdown for MiniMax-M3.1-Flash across the AA Intelligence Index — ArtificialAnlys · 2026-10-19
• [source] MiniMax-M3.1-Flash ranks #4 on Artificial Analysis Intelligence Index at $0.15/M input tokens — ArtificialAnlys · 2026-10-19

1 near-duplicate retellings: ArtificialAnlys