MiniMax launches M3.1 Flash: API available today, weights promised later

On October 19, 2026, MiniMax announced the launch of M3.1 Flash. The model is available through the company’s API, the M Plan subscription, and MiniMax Code; the MiniMax announcement describes the release of the weights as a future step but gives no exact date. For now, you can try the model through the company’s services; self-hosting will have to wait until the weights are released.

MiniMax lists prices of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. The company also claims that the total cost of agentic tasks with M3.1 Flash is 70% lower than with the previous version. Actual costs depend on the workload profile, including how much context can be reused and how many model calls are made.

Results reported by MiniMax

For coding, the company reports 73.7% on DeepSWE v1.1. For computer-use tests, it lists 61.7 on OSWorld V2, which evaluates task completion on a computer and across different applications, and 29.6 on Agents’ Last Exam (ALE), which evaluates the completion of professional tasks. For CyberGym, MiniMax reports a score of 82% in vulnerability discovery and verification. These results were published by MiniMax itself.

MiniMax also says that generation speed in ordinary use can reach 200 tokens per second, with median P50 speed exceeding 170 tokens per second. To compare models under your own workload, it is useful to measure time to first token, output speed, and the full cost of a completed task separately: a single speed metric does not describe the overall workflow.

How to test the model with your workload

Before moving tasks over, compare M3.1 Flash with your current model on a small set of real requests. For coding, keep the repository, instructions, and acceptance criteria the same; for an agent, keep the tools, permissions, and step limit consistent. Track task success, the amount of manual correction, repeat calls, and tokens per completed task. This lets the team assess the cost of the result, not just the price per token.

If your workflow requires local deployment, wait for the weights to be published and review their license, format, and infrastructure requirements once MiniMax makes them public. As of the announcement, the company confirmed access through its API, M Plan, and MiniMax Code, and said the weights would be released later.

Existing coverage includes articles about the release of DeepSeek V4.1 Flash and its web-search API. This article covers a separate MiniMax launch and the practical question of comparing the cost of agent calls and coding tasks.