MiniMax M3.1 Flash: 42 on Intelligence Index, about $0.22 per task, open weights planned for Oct 23
According to a report by The Decoder on October 19, 2026, MiniMax released the reasoning model M3.1 Flash; Artificial Analysis gives it an intelligence index score of 42, 13 points above the previous M3, at about $0.22 per task, with a 1 million token context and text and image input, and a planned open-weight release on October 23.
What changed
According to a report by The Decoder on October 19, 2026, MiniMax released the reasoning model M3.1 Flash. Citing Artificial Analysis data, the report says M3.1 Flash scores 42 on the intelligence index, 13 points above the previous M3, at less than half the latter's cost per task, about $0.22; it is level with GLM-5.3-Flash and slightly above Gemini 3.8 Flash's 41, and trails Deepseek V4.1 Flash 1017's 45. The report says the biggest gains are in complex reasoning: 48% on Humanity's Last Exam, above Deepseek and GLM; 22% on CritPt, a research-grade physics reasoning benchmark, ahead of GLM-5.3-Flash's 15% and Gemini 3.8 Flash's 18%, and slightly below Deepseek's 24%; and 56% correct answers on AA-Omniscience factual accuracy, far above GLM-5.3-Flash's 28% and Deepseek's 43%. The report also notes that, according to Artificial Analysis, its weakness is a 93% hallucination rate, attributed to the model's tendency to answer every question rather than decline when uncertain. The model supports up to a 1 million token context and accepts text and image input, and MiniMax says it runs three times faster than M3. The reported API pricing is $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached input, with the API already live on the MiniMax platform and accompanied by a desktop app called MiniMax Code. The report says MiniMax plans to release the model as open weights on October 23, while Deepseek V4.1 Flash 1017's weights are already on Hugging Face. The benchmarks, costs and speed figures above are third-party or vendor data relayed by the source.
What it means for developers
The direct implication for developers is that the combination of low cost, high throughput and high hallucination may change routing strategy rather than replace an established primary model outright. The source reports a 93% hallucination rate and a tendency to answer even when uncertain, which means that in factual question answering without retrieval support, and in regulatory or medical scenarios, treating its output as final fact carries relatively high risk; but the relatively low cost per task and lower output token consumption give it room to substitute in pipelines with a verification loop (draft generation, candidate expansion, bulk classification, first-draft code completion). The open-weight timing is also worth watching: the source says a release is planned for October 23, but open weights do not equal an open-source license, and downloadable weights do not automatically grant commercial use, redistribution or post-fine-tuning rights; the actual terms must be checked against the license text of the release repository. Note the version naming and hosting status: as of the catalog check for this article, the exact version MiniMax M3.1 Flash does not appear in the live KimiSeek catalog, so the API availability and pricing described in the source refer to MiniMax's official channel, not this site; the DeepSeek V4.1 Flash 1017 and GLM-5.3 Flash present in the catalog are different versions and are not evidence that this version is available. The reported $0.15 input and $0.60 output figures are as of the October 19, 2026 reporting, and prices can change at any time; actual billing should follow the official pricing page.
Integration and evaluation
The following is advice and conditional analysis, not facts stated by the source. First decision: whether to introduce M3.1 Flash into a high-throughput candidate generation layer. Based on the lower cost per task and lower output token consumption described in the source, if your workload is short-output and high-concurrency (for example query rewriting, tag expansion, log categorization), you could first run a shadow-traffic comparison, sending the same requests to both your current model and M3.1 Flash, with evaluation metrics set to factual spot-check pass rate and cost per thousand requests; when the spot-check pass rate falls below your availability threshold (for example more than 5 percentage points below your current model), keep it only for candidate expansion and not for final output. The precondition is that the business allows downstream human or rule-based verification. Second decision: whether to use the 1 million token context and text and image input to rework long-document pipelines. Based on the context limit and text-plus-image input described in the source, if your workload involves long reports or multi-page scans, consider feeding the whole document instead of chunk-and-concatenate, with evaluation metrics set to answer citation grounding accuracy and end-to-end latency; when the hallucination rate worsens further under long context and the spot-check citation error rate rises, fall back to chunking plus retrieval. Third decision: deployment and compliance after the open-weight release. Based on the October 23 open-weight plan described in the source, if you want to self-host, the precondition is that the license terms permit your use case; the evaluation metric is re-testing factual accuracy and refusal rate on your own corpus, and using that to decide whether to add a mandatory "decline when information is insufficient" prompt or validator at the inference layer, because the source explicitly notes the model's tendency not to decline.
Limits and open questions
The source is a third-party report dated October 19, 2026; the benchmark scores, cost per task, hallucination rate and speed multiple are data relayed from Artificial Analysis or MiniMax and have not been independently reproduced. The definition and test conditions of the 93% hallucination rate are not fully given in the source and should not be directly extrapolated to your business corpus. The open weights are a plan, not an accomplished fact, and open weights do not equal an open-source license; commercial and redistribution rights need to be verified against the actual license. To be clear: as of the catalog check for this article, the exact version MiniMax M3.1 Flash does not appear in the live KimiSeek catalog, and the API availability and pricing described in the source refer to MiniMax's official channel; other models in the catalog positioned as Flash are different versions and are not grounds for concluding that this version is available. The pricing reported in the source reflects the situation at that time and may have changed; the official pricing page should be taken as authoritative.
Illustrations from the original report
