Key Takeaways
- • MiniMax's M3.1 Flash scored 42 on Artificial Analysis's Intelligence Index, up 13 points from its predecessor M3's 29, while cutting the average cost per Intelligence Index task from roughly $0.51 to $0.22, a reduction of nearly 57%.
- • Artificial Analysis found M3.1 Flash correctly answered 56% of AA-Omniscience factual questions, but produced a 93% hallucination rate, attempting nearly every question and overwhelmingly giving incorrect answers rather than admitting uncertainty.
- • MiniMax prices M3.1 Flash at $0.15 per million input tokens and $0.60 per million output tokens, a total of $0.75 per million tokens in a 1:1 workload, placing it among the cheapest frontier-class models listed.
- • The model is competitive with DeepSeek V4.1 Flash-1017 (scoring 45 at roughly $0.26 per task) but trails the strongest Anthropic and OpenAI models on the overall index, according to the independent results.
- • MiniMax reports 62.2 on Automation-Bench and 76.2 on BankerToolBench, tying Claude Opus 5 on the latter, though these are vendor-reported benchmark results rather than production deployments.
Summary
Chinese AI company MiniMax has released M3.1 Flash, a model designed for complex enterprise workflows, software development and computer-use tasks. MiniMax says the model runs nearly three times faster than its predecessor and reduces overall usage costs for AI agents by 70%. Independent testing by Artificial Analysis provides a more detailed picture. The firm awarded M3.1 Flash a score of 42 on its Intelligence Index at maximum reasoning effort, compared with 29 for M3, with the average cost per task falling from about $0.51 to $0.22. That places the model on par with Z.ai's GLM-5.3 Flash (42 at about $0.28 per task) and just ahead of Google's Gemini 3.8 Flash (41 at about $1.24 per task). DeepSeek V4.1 Flash-1017 leads the group at 45 for about $0.26 per task. Artificial Analysis also noted that M3.1 Flash remains behind the most capable frontier systems from Anthropic and OpenAI on the overall index.
On specific benchmarks, Artificial Analysis reported that M3.1 Flash scored 48% on Humanity's Last Exam, up from 39% for M3, and 22% on the CritPt physics reasoning benchmark, up from 4%. The model also used fewer output tokens during the Intelligence Index evaluation, approximately 189 million compared with about 324 million for GLM-5.3 Flash and 334 million for Gemini 3.8 Flash, which matters because developers pay for generated tokens, including intermediate reasoning. In its own announcement, MiniMax reported 62.2 on Automation-Bench v1.0.6, narrowly ahead of DeepSeek V4.1 Flash-1017's 60.5, and 76.2 on BankerToolBench, tying Claude Opus 5. Its coding result of 73.7% on DeepSWE v1.1 trails Claude Opus 5 (74%) and DeepSeek V4.1 Flash-1017 (77%). These are benchmark results rather than documented production deployments, and differing evaluation settings limit direct comparison.
The main caveat concerns factual reliability. On AA-Omniscience, M3.1 Flash correctly answered 56% of questions, versus 43% for DeepSeek V4.1 Flash-1017 and 28% for GLM-5.3 Flash. However, Artificial Analysis found that the model attempted nearly every question, producing a 93% hallucination rate under the benchmark's definition, which measures the share of non-correct responses that were outright wrong rather than partial answers or refusals. That behavior yielded an overall AA-Omniscience score of just 16. Artificial Analysis's findings suggest enterprises deploying the model for autonomous record updates, financial document analysis or customer research may need verification layers, retrieval systems, guardrails or human review.
MiniMax offers M3.1 Flash through its first-party API, its M Plan subscription and its MiniMax Code desktop application, priced at $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015 per million. According to Artificial Analysis, the model supports text and image inputs, text outputs and a one-million-token context window. MiniMax claims output speeds above 170 tokens per second median and attributes part of the efficiency gains to increased sparsity. The weights remain proprietary at launch, with MiniMax planning to release them on October 23.
On specific benchmarks, Artificial Analysis reported that M3.1 Flash scored 48% on Humanity's Last Exam, up from 39% for M3, and 22% on the CritPt physics reasoning benchmark, up from 4%. The model also used fewer output tokens during the Intelligence Index evaluation, approximately 189 million compared with about 324 million for GLM-5.3 Flash and 334 million for Gemini 3.8 Flash, which matters because developers pay for generated tokens, including intermediate reasoning. In its own announcement, MiniMax reported 62.2 on Automation-Bench v1.0.6, narrowly ahead of DeepSeek V4.1 Flash-1017's 60.5, and 76.2 on BankerToolBench, tying Claude Opus 5. Its coding result of 73.7% on DeepSWE v1.1 trails Claude Opus 5 (74%) and DeepSeek V4.1 Flash-1017 (77%). These are benchmark results rather than documented production deployments, and differing evaluation settings limit direct comparison.
The main caveat concerns factual reliability. On AA-Omniscience, M3.1 Flash correctly answered 56% of questions, versus 43% for DeepSeek V4.1 Flash-1017 and 28% for GLM-5.3 Flash. However, Artificial Analysis found that the model attempted nearly every question, producing a 93% hallucination rate under the benchmark's definition, which measures the share of non-correct responses that were outright wrong rather than partial answers or refusals. That behavior yielded an overall AA-Omniscience score of just 16. Artificial Analysis's findings suggest enterprises deploying the model for autonomous record updates, financial document analysis or customer research may need verification layers, retrieval systems, guardrails or human review.
MiniMax offers M3.1 Flash through its first-party API, its M Plan subscription and its MiniMax Code desktop application, priced at $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015 per million. According to Artificial Analysis, the model supports text and image inputs, text outputs and a one-million-token context window. MiniMax claims output speeds above 170 tokens per second median and attributes part of the efficiency gains to increased sparsity. The weights remain proprietary at launch, with MiniMax planning to release them on October 23.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.