MiniMax unveils M3.1 flash: 428b moe model at $0.15 per million input tokens

MiniMax has launched M3.1 Flash, a multimodal mixture-of-experts model with 428B total parameters and 23B active parameters, available via its API at $0.15 per million input tokens. The model supports text, image, and video inputs, with a context window of 1 million tokens. It is priced at half the rate of its standard M3 counterpart, which costs $0.30 per million input tokens. Reasoning is enabled by default and cannot be disabled, with developers able to choose from five effort levels. The model’s weights are set to be released on October 23, 2026. M3.1 Flash outperforms several competitors on benchmark tests, including Automation-Bench v1.0.6 with a score of 62.2, surpassing GPT-5.6 Sol at 45.8. It also leads in visual reasoning tasks, achieving 91.0 on BabyVision with tools. However, coding benchmarks show mixed results, with DeepSeek-V4.1-Flash-1017 scoring higher at 77.0. MiniMax claims the model offers near-frontier agentic performance at Flash-tier costs, though independent validation remains pending. The model is accessible through platform.minimax.io, with Anthropic and OpenAI-compatible APIs. Pricing includes $0.60 per million output tokens and $0.015 for cache reads. MiniMax has not disclosed the license terms or self-hosting details, which are expected to be available by October 23, 2026.