Technology

MiniMax M3.1 Flash ties GLM-5.3-Flash and edges past Gemini on Artificial Analysis index

Cost per task undercuts Deepseek at $0.22, open weights land October 23 as China's price race shows no floor

Images

Image description Image description the-decoder.com
MiniMax-M3.1-Flash jumps from 29 to 42 points on the Artificial Analysis Intelligence Index compared to M3, while cutting cost per task sharply. | Image: AA MiniMax-M3.1-Flash jumps from 29 to 42 points on the Artificial Analysis Intelligence Index compared to M3, while cutting cost per task sharply. | Image: AA AA
On DeepSWE v1.1, MiniMax-M3.1-Flash scores up to 73.7 percent at under $0.30 per task. | Image: MiniMax On DeepSWE v1.1, MiniMax-M3.1-Flash scores up to 73.7 percent at under $0.30 per task. | Image: MiniMax MiniMax

MiniMax has released M3.1 Flash, a reasoning model that scores 42 points on the Artificial Analysis Intelligence Index — thirteen points above its predecessor at less than half the cost per task. The score ties GLM-5.3-Flash, edges past Gemini 3.8 Flash at 41, and trails Deepseek V4.1 Flash at 45.

API pricing is $0.15 per million input tokens and $0.60 per million output tokens, which works out to about $0.22 per task on the benchmark suite — slightly under Deepseek's $0.26. The margin comes from token discipline: MiniMax used roughly 189 million output tokens across the benchmarks against Deepseek's 224 million, and the company says the new model runs three times faster than the last. It accepts text and images, handles up to one million tokens of context, and scores up to 73.7 percent on the DeepSWE coding benchmark at under $0.30 per task.

The benchmark picture is mixed but competitive. On Humanity's Last Exam, M3.1 Flash scores 48 percent, ahead of both Deepseek and GLM. On CritPt, a test of research-level physics reasoning across 71 problems, it manages 22 percent against Deepseek's 24. On AA-Omniscience, a factual-accuracy test, it answers 56 percent of questions correctly against GLM's 28.

Both MiniMax and Deepseek are releasing their models as open weights — Deepseek is already on Hugging Face, and MiniMax follows on October 23. Capability that Western laboratories sell by the token is being posted for free download. The Decoder, which reported the figures, frames the release as further evidence that China's AI price race has no floor in sight. Laboratories that must pay down enormous datacenter commitments need the margin that inference pricing provides; laboratories without those commitments can cut prices and let competitors worry about the arithmetic.

Artificial Analysis puts M3.1 Flash's hallucination rate at 93 percent, attributing it to a single habit: the model attempts every question rather than declining when uncertain.