MiniMax M3.1 Flash shows China's AI price race has no floor in sight

MiniMax has released M3.1 Flash, a new reasoning model that lifts the company's position in the crowded low-cost end of the Chinese model market. On the Artificial Analysis Intelligence Index, M3.1 Flash scores 42 points, a 13-point gain over its predecessor M3, while cutting cost per task to less than half. The API is already live at platform.minimax.io, and MiniMax says it plans to publish the model as open weights on October 23. The company also ships a desktop application called MiniMax Code.
The headline number is the cost. Benchmarking by Artificial Analysis puts M3.1 Flash at about $0.22 per task, which the model reaches largely by emitting fewer output tokens than its competitors rather than by pricing alone. MiniMax's API list prices are $0.15 per million input tokens and $0.60 per million output tokens, with cached input billed at $0.015 per million. The model accepts both text and images and handles up to one million tokens of context. MiniMax also claims it runs three times faster than M3.
On the Intelligence Index, the 42-point score ties GLM-5.3-Flash, edges past Gemini 3.8 Flash at 41, and trails Deepseek V4.1 Flash 1017 at 45. The largest improvement over M3 came in complex reasoning. On Humanity's Last Exam, M3.1 Flash reaches 48 percent, ahead of both Deepseek and GLM. On CritPt, a benchmark of research-level physics reasoning built from 71 complex problems, it scores 22 percent, above GLM-5.3-Flash at 15 percent and Gemini 3.8 Flash at 18 percent, and just behind Deepseek at 24 percent. On AA-Omniscience, which measures factual accuracy, it answers 56 percent of questions correctly, compared with 28 percent for GLM-5.3-Flash and 43 percent for Deepseek.
Coding performance is reported through DeepSWE v1.1, where MiniMax-M3.1-Flash scores up to 73.7 percent at under $0.30 per task. The source describes this as matching Deepseek V4.1 Flash 1017 on both performance and price.
One clear weak point sits alongside the accuracy gains. Artificial Analysis reports a 93 percent hallucination rate for the model, attributing it to a tendency to attempt every test question rather than decline when uncertain. That is a meaningful trade-off for anyone deploying the model in open-ended settings: a model that almost never abstains will produce confident answers in cases where it should defer to a human or a retrieval step. In practice, teams building on M3.1 Flash should expect to add their own guardrails, such as retrieval grounding, confidence checks or refusal prompting, rather than relying on the model to signal uncertainty on its own.
Against Deepseek V4.1 Flash 1017, neither model dominates. Deepseek leads on the overall Intelligence Index, 45 versus 42, and holds a slight edge on CritPt at 24 percent versus 22 percent. MiniMax wins on Humanity's Last Exam, 48 percent versus 43 percent, and on AA-Omniscience accuracy, 56 versus 43 percent, and comes in slightly cheaper per task at $0.22 versus $0.26. The cost gap comes from token efficiency: M3.1 Flash needed roughly 189 million output tokens across the benchmark run, compared with about 224 million for Deepseek. Both models are headed for open-weight distribution; Deepseek's is already available on Hugging Face, while MiniMax's is scheduled for October 23.
A one-million-token context window with multimodal input at these prices is notable for product teams weighing long-document or agent workflows, where prompt size and repeated calls tend to dominate spend. Cached input at $0.015 per million tokens makes repeated long prompts meaningfully cheaper on paper, though actual savings depend on how well a given application reuses context. The lower output-token count is arguably the more durable lever, since it lowers cost on every call rather than only on cache hits, and it can also reduce latency in multi-step pipelines.
Why it matters: Developers choosing a low-cost reasoning model now face a genuinely close contest rather than a clear winner, with MiniMax and Deepseek trading leads across benchmarks at nearly identical prices. The 93 percent hallucination rate is the item most likely to shape deployment decisions, since it argues for architectures that verify or constrain outputs rather than trusting the model to decline. Product teams should also note the October 23 open-weights date, which, if it holds, would let them self-host rather than depend on the hosted API.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
MiniMax's M3.1 Flash jumps 13 points over its predecessor to 42 on the Artificial Analysis Intelligence Index, at half the cost. It also hits 73.7 percent on the DeepSWE v1.1 coding benchmark for under $0.30 per task, matching Deepseek V4.1 Flash 1017 on both performance and price. The article MiniMax M3.1 Flash shows China's AI price race has no floor in sight appeared first on The Decoder .