In brief
- M3.1 Flash scored 42 on the Artificial Analysis Intelligence Index, compared with 29 for the previous M3 model.
- In a comparison reported by The Decoder, the model costs around $0.22 per task; DeepSeek V4.1 Flash 1017 costs $0.26.
- The model accepts text and images and supports a context window of up to one million tokens.
- MiniMax plans to release the model’s open weights on October 23.
MiniMax announced its reasoning model, M3.1 Flash, on October 19, 2026. The model’s API is now available, and MiniMax said it plans to release an open-weights version on October 23. According to information reported by The Decoder, M3.1 Flash can process text and image inputs and supports a context window of up to one million tokens.
How did it perform in tests?
M3.1 Flash scored 42 on the Artificial Analysis Intelligence Index, up 13 points from the previous M3 model’s score of 29. In the same comparison, DeepSeek V4.1 Flash 1017 led with 45 points, while Gemini 3.8 Flash scored 41. The Decoder reports that the MiniMax model outperformed competitors on some reasoning and accuracy benchmarks.
M3.1 Flash scored 48% on the “Humanity’s Last Exam” benchmark, compared with DeepSeek’s reported 43%. On CritPt, which measures research-level reasoning in physics, M3.1 Flash scored 22% and DeepSeek scored 24%. The reported accuracy rate on the AA-Omniscience benchmark was 56% for the MiniMax model and 43% for DeepSeek. Results vary according to the capabilities measured by each benchmark; no single score describes every use case.
API pricing and access options
According to The Decoder, API pricing is $0.15 per million input tokens and $0.60 per million output tokens. Cached inputs cost $0.015 per million tokens. In the comparison reported in the article, the model costs around $0.22 per task, while the same metric for DeepSeek V4.1 Flash 1017 is $0.26.
MiniMax says M3.1 Flash runs three times faster than M3; this performance claim is based on the company’s announcement. The article also notes that the model’s lower output-token usage helps reduce the cost per task. The MiniMax Code desktop app is another access option. The announced release date for the open-weights version is October 23; the source text does not specify its licensing terms.
Impact on architecture and visualization workflows
Support for text and image inputs, along with a large context window, makes the model suitable for experimentation with tasks such as project documentation, visual references, and text-based research. However, the sources do not indicate that M3.1 Flash has integrations for architecture software or direct 3D modeling and rendering capabilities. It is therefore more accurate to view it as an assistant for language and image analysis than as a replacement for production tools.
For firms and visualization teams, API costs should be calculated against usage volume; separate charges for input and output tokens can affect budgets. A one-million-token context window may make it possible to work with large document collections, but the model’s accuracy limitations should also be taken into account. The 93% hallucination rate reported in the Artificial Analysis evaluation cited by The Decoder underscores the need to independently verify project decisions and technical information.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
What stands out about M3.1 Flash is not just a single benchmark score, but its combination of strong results in some comparisons and a low cost per task. For architecture and visualization teams in Turkey, this could be a reason to test it in a controlled way, particularly for research workflows involving text and visual references. However, the sources do not verify any integration specifically for architecture software.
Before making a decision, firms should measure API usage against their own workloads and clarify the licensing and open-weights terms. The reported hallucination rate also shows why model outputs should not become project inputs without technical verification. Low cost is no substitute for checking accuracy.
Frequently asked questions
When was MiniMax M3.1 Flash released?
MiniMax announced the model and made its API available on October 19, 2026. The open-weights version is planned for release on October 23.
How much does the MiniMax M3.1 Flash API cost?
Pricing is $0.15 per million input tokens and $0.60 per million output tokens. Cached inputs cost $0.015 per million tokens.
Does M3.1 Flash support image inputs?
Yes. According to the source, the model accepts both text and image inputs and supports a context window of up to one million tokens.
