MiniMax launches M3.1 Flash a cheap agentic AI model with a 1 million token context

MiniMax has released its newest model, MiniMax M3.1 Flash, on its API platform. The model is priced at just $0.15 per million input tokens and $0.60 per million output tokens, which is half the cost of MiniMax's earlier M3 model.

MiniMax M3.1 Flash is a multimodal model, meaning it can process text, images and video. It is built using a technique called Mixture of Experts (MoE), where the model has 428 billion parameters total but only activates about 23 billion at a time. This means it can handle complex tasks without needing as much computing power as a standard model. The model also features a one million token context window — large enough to process the equivalent of thousands of pages of text in a single request.

Built-in reasoning, always on

One major change in M3.1 Flash is that reasoning cannot be turned off. When the model thinks through a problem, that thinking appears in a separate field in the response. Developers can choose from five effort levels, ranging from Low to Max, with MiniMax recommending Max as the default setting.

This always-on reasoning means every request uses more tokens than it would if the model simply responded immediately. Developers who previously used reasoning-disabled shortcuts will need to update their integrations, as those requests now return errors.

Strong on agent tasks, weaker on coding

On agentic benchmark tests — measures of how well a model can handle multi-step office and automation work — MiniMax M3.1 Flash leads the pack at 62.2 on Automation-Bench, compared with 50.3 for Anthropic's Opus 5 and 45.8 for OpenAI's GPT-6 Sol. It also topped the LOCA-Bench at 60.3.

However, on harder coding benchmarks the model sits in the middle of the field. On Terminal Bench 4.0, it scored 38.1, behind Opus 5 at 51.8. On DeepSWE v1.1 it scored 73.7, just below DeepSeek V4.1 Flash 1017 at 77.0.

MiniMax reports that the model runs at roughly 150 tokens per second and claims it is three times faster than the original M3 model.

What is missing

All of the benchmark scores are reported by MiniMax and have not been independently verified by third parties. A comparison with Kimi K3's published card shows some alignment on DeepSWE and BabyVision scores, but MiniMax's Automation-Bench chart lists 62.8 while its table says 62.2, suggesting minor inconsistencies in reporting.

The company also has not yet disclosed the self-hosting hardware requirements or the license for the model weights, which are expected to arrive on October 23, 2026.

For developers looking for a cost-effective agent model with a very long context window, M3.1 Flash offers strong automation benchmarks at a fraction of the price of competing models.