MiniMax releases MiniMax-M3.1-Flash with a 1M-token context window

MiniMax released MiniMax-M3.1-Flash on 19 October 2026, through the MiniMax API, the M Plan subscription and MiniMax Code. It keeps a context window of 1,000,000 tokens, and it accepts text, images and video, so an agent session can carry a long document and look at pictures or clips in the same request.

Thinking depth is a setting MiniMax calls effort: low, medium, high, xhigh or max. A higher setting makes the model think longer before it answers, which uses more output tokens and adds latency. If the request leaves effort unset, the default is max. Thinking stays on, and MiniMax says a request that tries to turn it off is rejected. On the OpenAI-style API the reasoning comes back in a separate field from the answer.

MiniMax says Space Bunny, which processed more than 200 trillion tokens while running without its name attached, was an early version of this model.

The comparison table and the two cost charts are MiniMax’s. The table prints 62.2 on Automation-Bench v1.0.6 while the Automation-Bench cost chart prints 62.8% at max effort, and the post’s claim of usage cost 70% lower than the predecessor is not the list price: the pricing page sets input at half MiniMax-M3’s standard rate on inputs up to 512k tokens.

Key numbers

• Input $0.15, output $0.60 and cached input $0.015, per million tokens, on the pricing page
• Context window of 1,000,000 tokens
• Thinking depth from low to max, with max as the default

The release notes call it an open-weight model. The research post says the weights will be released later, and does not give a date. MiniMax also says a Pro model will arrive later, again with no date.