Overview
- MiniMax officially released M3.1 Flash on Monday and made it available through its API, M Plan, and MiniMax Code with published per‑token rates for input, output, and cached input.
- The company says the model runs inference nearly three times faster than M3, supports a one‑million‑token context window, and cuts overall agent usage costs by 70 percent using new rates of $0.15 input, $0.60 output, and $0.015 cached input per 1M tokens.
- Independent tests reported by The New Stack and Artificial Analysis found smaller cost gains and accuracy problems, with one sampled workload showing a 57 percent cost drop versus MiniMax’s 70 percent claim and factual‑question checks finding 56 percent correct answers and a high rate of incorrect responses when the model attempted questions.
- Key items remain unresolved: MiniMax has not released model weights or a licensing text, security and benchmark claims come from company‑run evaluations, and press reporting says weights may be released on Oct. 23 but that timing is unconfirmed by MiniMax.
- Practical impact will vary because agent task cost depends on how many tokens an agent consumes in loops and tool calls, so the large context and lower rates could change developer workflows but also raise the need for independent validation and human verification for knowledge‑heavy or security uses.