MiniMax, the Hong Kong-listed Chinese AI company, has launched M3.1 Flash, a reasoning model priced at 15 cents per million input tokens. The company says agentic tasks cost 70% less to run than on its previous model, M3, and that inference is close to three times as fast. Independent testing by Artificial Analysis, reported by The Decoder, supports part of that story, notably a large cost reduction per task, but also flags a very high hallucination rate. The weights are due on October 23, according to The Decoder.

What MiniMax announced

M3.1 Flash is available now through the MiniMax API, the company’s M Plan subscription and its MiniMax Code desktop app. Output costs 60 cents per million tokens, and cached input costs 1.5 cents. MiniMax says peak output at everyday load tops 200 tokens per second.

According to SiliconANGLE, the model is barely larger than M3: MiniMax’s developer guide for the preview lists 428 billion parameters in total, and the company credits the speed to more aggressive sparsity at inference time, meaning fewer of those parameters are used for each token. Reasoning is always on, and developers pick one of five effort levels from Low to Max. The prompt window reaches 1 million tokens. SiliconANGLE reports text, image and video input, while The Decoder lists text and images, so check the documentation for the modalities your use case needs.

MiniMax also said it will publish the Flash weights later and that a more capable Pro model is coming. The Decoder gives the open-weight date as October 23.

The numbers, and who produced them

Most headline scores come from MiniMax itself. It reported 61.7 on OSWorld V2, a test of computer operation, saying no Chinese model has done better; 76.2 on BankerToolBench, a banking workflow test; and 82% on CyberGym, a Berkeley-built benchmark of whether agents can reproduce real software vulnerabilities. On DeepSWE v1.1, a software engineering test, M3.1 Flash reached 73.7%. DeepSeek said its updated V4.1-Flash resolved 77% on the same test. SiliconANGLE notes that neither result has been independently verified.

Third-party figures give a more measured picture. Artificial Analysis scored M3.1 Flash at 42 on its Intelligence Index, up from 29 for M3. That ties it with GLM-5.3-Flash, edges Gemini 3.8 Flash (41) and trails DeepSeek V4.1 Flash 1017 (45). The cost to run the index fell to about $0.22 per task, against roughly $0.26 for DeepSeek, because MiniMax used about 189 million output tokens across the benchmark versus 224 million for DeepSeek.

The mixed results continue elsewhere. M3.1 Flash scored 48% on Humanity’s Last Exam, ahead of DeepSeek’s 43%, and answered 56% of AA-Omniscience questions correctly, against 43% for DeepSeek. On CritPt, a research-level physics test, it reached 22% against DeepSeek’s 24%. The weak spot is a 93% hallucination rate on Artificial Analysis’s measure, because the model tries to answer every question rather than declining when unsure.

The price war and the stealth model

M3.1 Flash lines up directly with DeepSeek’s V4.1-Flash, which charges the same 15 cents per million input tokens off-peak. SiliconANGLE reports that MiniMax’s shares fell after the original V4.1-Flash launch in September, citing Bloomberg. MiniMax has traded in Hong Kong since a $619 million IPO in January.

The launch also resolves a small mystery. An anonymous model called Space Bunny appeared on OpenRouter’s routing service in late September, and community testers linked it to MiniMax by probing its tokenizer. The company confirmed that Space Bunny was an early M3.1 Flash build and said more than 200 trillion tokens passed through it while its maker was unnamed.

What this means for buyers

A lower per-token price does not automatically lower a bill. Agentic workloads multiply tokens through repeated tool calls and long reasoning traces, so total cost depends on how many tokens a model needs to finish a task. The Artificial Analysis figure of cost per task is therefore more useful than the headline rate, and it is the right unit to compare when cutting your API bill. Our explainer on how AI pricing works covers the basics.

Teams should also weigh the hallucination figure. A fast, cheap agent that confidently answers unanswerable questions can create expensive errors downstream, and federal restrictions on some Chinese models in the United States, including a new DHS order on DeepSeek, may matter for regulated buyers.

What to watch next

  • The October 23 weight release: license terms, file sizes and hardware requirements.
  • Independent reproduction of the OSWorld V2 and CyberGym claims.
  • A response from DeepSeek, Zhipu and others on price for agentic workloads.
  • The arrival of the Pro model MiniMax says is coming.
  • Whether the hallucination rate improves in later checkpoints.

Key takeaways

  • M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens.
  • MiniMax claims 70% lower agent task costs; Artificial Analysis measured about $0.22 per task.
  • On Artificial Analysis’s index it scores 42, behind DeepSeek V4.1 Flash 1017 at 45.
  • Its 93% hallucination rate means it rarely declines to answer when uncertain.
  • Open weights are reported for October 23.

Frequently asked questions

Is M3.1 Flash open source?

Not yet. MiniMax says it will release the weights later, and The Decoder reports October 23. Check the license when it appears.

Is it better than DeepSeek V4.1 Flash?

Neither wins across the board. DeepSeek leads the Artificial Analysis index, 45 to 42, while MiniMax is ahead on Humanity’s Last Exam and factual accuracy and slightly cheaper per task.

Where was Space Bunny from?

MiniMax confirmed that the anonymous Space Bunny model on OpenRouter was an early build of M3.1 Flash.