MiniMax has launched M3.1 Flash, an AI model aimed at coding and agentic work that the company says can run at more than 200 tokens per second while cutting inference costs by 70%. It is available through MiniMax’s API, M Plan and MiniMax Code, giving users a new model to try rather than another “coming soon” teaser.
MiniMax Watch analysis
What happened
MiniMax describes M3.1 Flash as built for agentic tasks, coding, computer use and cybersecurity. The company says it has improved performance on AutomationBench-AA, DeepSWE and CyberGym, and highlights a one-million-token context window and tunable thinking depth.
Those performance and cost figures are MiniMax’s claims. The announcement does not include the benchmark scores needed to compare the model’s results directly, and says model weights will be released later.
Why it matters
The useful proposition is speed and cost together: if M3.1 Flash can handle multi-step coding or automation work while using less inference spend, that could make more frequent agent runs practical. Availability across an API and MiniMax’s own products also means users have immediate ways to test it.
But “70% cheaper” is not a universal price tag. The saving depends on what is being compared, the workload and the quality of the result. A fast model that needs extra retries can make the spreadsheet less cheerful.
Our read
This is a substantive model launch, not just a name and a promise. MiniMax has put specific claims on the table, but the missing benchmark figures leave the most important comparison for users and independent evaluators to settle. Try it against a real task and a known alternative before rebuilding a workflow around the headline numbers.
What to watch
- Whether MiniMax publishes benchmark scores and enough detail to reproduce the comparisons.
- How the claimed speed and cost savings hold up on real coding and agent workloads.
- When the model weights arrive, and what hardware and licence terms they come with.
Discussion spark: Would a 70% reduction in inference cost persuade you to switch models, or would you need independent task results first?
Sources and evidence
- MiniMax M3.1 Flash: Redefining Agent Performance and Cost – MiniMax Research | MiniMax (19 October 2026, 12:11 UTC)
not affiliated with or endorsed by MiniMax