Technology · NEWS EXPLAINER
MiniMax launches M3.1 Flash: API prices, 1M context and the claims to test
The new model is live through MiniMax’s API, M Plan and MiniMax Code. Its speed, cost and benchmark figures are company-reported and still need workload-specific verification.
MiniMax launched M3.1 Flash on October 19, adding a new model for agentic, coding, computer-use and cybersecurity work. Reuters independently confirmed the launch and summarized the company’s claim that online inference is nearly three times faster while usage costs are 70% lower.
The official release and API documentation add details that matter to developers: one-million-token context, tunable reasoning effort, multimodal inputs and flat token prices. The article below separates what is available now from benchmark and performance claims that have not been independently reproduced in the sources reviewed.
What launched on October 19
MiniMax released M3.1 Flash on October 19. Reuters independently reported the launch, while the company’s release page says the model is available through the MiniMax API, M Plan and MiniMax Code. The model weights were not released with the announcement; MiniMax says they will come later.
What developers can use now
MiniMax’s current API documentation lists the model identifier as MiniMax-M3.1-Flash and supports both Anthropic-compatible and OpenAI-compatible endpoints. The documentation says the model accepts text, image and video inputs, supports a one-million-token context window and offers low, medium, high, xhigh and max reasoning-effort settings. Thinking cannot be disabled; developers can reduce effort to trade depth for latency and token use.
The published API prices
MiniMax lists flat rates of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached-input tokens. The company says overall agent-task usage cost is 70% lower than its predecessor. Those are launch prices and a vendor comparison, not a guarantee of the bill for every workload; prompts, outputs, caching and tool loops still determine total usage.
Speed and benchmark claims need context
The release says online inference is nearly three times faster, with everyday output exceeding 200 tokens per second and median speed above 170. It also publishes scores for DeepSWE, OSWorld V2, Agents’ Last Exam and CyberGym. Reuters reports the headline speed and cost claims, but neither reviewed source provides an independent replication of those benchmarks. Readers should treat them as company-reported launch results until third-party testing uses comparable settings.
A practical evaluation checklist
Before moving production traffic, teams should test their own prompts at the intended effort level, record end-to-end latency rather than only token speed, measure tool-call success and compare total token use with the current model. Security teams should separately validate any code-audit or cybersecurity task in an isolated environment. The launch establishes availability and pricing; it does not establish that every workflow will be faster, cheaper or more accurate.
