MiniMax M3.1 Flash cuts cost but still trails DeepSeek
MiniMax's new Flash model reaches 42 on the Artificial Analysis index at about $0.22 per task. A 93% hallucination rate complicates the price story.
MiniMax M3.1 Flash is a real price-performance step, but not a clean win. It scores 42 on the Artificial Analysis Intelligence Index at about $0.22 per task, yet DeepSeek V4.1 Flash 1017 still scores 45, and Artificial Analysis reportedly measures a 93% hallucination rate, which matters for production use.
The Weights Desk · 3 min read- MiniMax M3.1 Flash scores 42 on the Artificial Analysis Intelligence Index, up 13 points from M3's 29, at under half M3's cost per task.
- DeepSeek V4.1 Flash 1017 still leads the index at 45 and costs more per task, about $0.26 versus $0.22, per The Decoder's reporting.
- Artificial Analysis reportedly records a 93% hallucination rate: the model answers nearly every question instead of declining when unsure.
- The 73.7% DeepSWE v1.1 figure comes from MiniMax's own chart, with no independent run cited, so treat it as a vendor claim.
- MiniMax lists API prices of $0.15 per million input tokens and $0.60 per million output tokens; open weights are promised but the date is unconfirmed by MiniMax.
MiniMax M3.1 Flash moves the cheap end of the market, but the headline numbers need caveats before anyone swaps it into production. According to The Decoder, it scores 42 on the Artificial Analysis Intelligence Index, up 13 points from M3, at under half M3's cost per task. DeepSeek V4.1 Flash 1017 still scores 45. The most consequential figure is also the least flattering: a reported 93% hallucination rate.
The price gain is real but narrow
The cost advantage over DeepSeek is measurable but small. The Decoder reports about $0.22 per task for MiniMax M3.1 Flash against $0.26 for DeepSeek V4.1 Flash 1017, helped by roughly 189 million output tokens across the benchmark versus 224 million. MiniMax lists API prices of $0.15 per million input tokens and $0.60 per million output tokens. That is a gap of about 15%, not a collapse in price.
Accuracy scores hide a hallucination problem
M3.1 Flash scores 56% on AA-Omniscience, well above DeepSeek's 43%, yet Artificial Analysis reportedly puts its hallucination rate at 93%. Those figures are consistent: the model answers almost every question rather than abstaining, so it gets more right and also states more wrong answers with confidence. For retrieval, compliance or support workloads, the abstention behavior matters more than the headline accuracy number.
The coding number is a vendor claim
MiniMax reports 73.7% on DeepSWE v1.1, and its announcement states that figure without a cost per task. The under-$0.30 per-task figure and the comparison with DeepSeek come from The Decoder's reading of a MiniMax chart; the article links no independent DeepSWE run. Until a third party reproduces it with a described harness, the coding result is a lead worth testing, not a ranking.
What to watch before adopting it
The open-weights release is the next checkpoint. The Decoder reports October 23, while MiniMax's own post says only that weights will come later, so the date is unconfirmed. Once weights are out, teams can measure latency, abstention behavior and cost on their own traffic. MiniMax says inference is nearly three times faster without naming the baseline, so that claim also needs testing.
Verdict: a credible cost-cutting option for tolerant workloads such as drafting, triage and code assistance with human review. It is not yet a safe default for factual answers, because DeepSeek is stronger on the aggregate index and M3.1 Flash rarely declines to answer. Run your own evaluation, including an abstention test, before moving traffic.
- Is MiniMax M3.1 Flash better than DeepSeek V4.1 Flash 1017?
- Not overall. DeepSeek leads the Intelligence Index 45 to 42 and CritPt 24% to 22%. MiniMax leads Humanity's Last Exam (48% vs 43%) and AA-Omniscience (56% vs 43%), and is cheaper per task, per The Decoder.
- What does the 93% hallucination rate mean for deployment?
- Artificial Analysis reportedly finds the model attempts nearly every question rather than abstaining. A team using it for retrieval or factual answers would need its own grounding checks or abstention logic before trusting the accuracy score.
- How much does MiniMax M3.1 Flash cost?
- MiniMax lists $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. Artificial Analysis cost per task, as reported by The Decoder, is about $0.22.