MiniMax launches M3.1 Flash, says agent costs drop 70%

Summary

• MiniMax released M3.1 Flash on October 19, 2026, targeting coding, computer use and cybersecurity agents.
• The company prices it at $0.15 per million input tokens and $0.60 per million output tokens.
• Cheaper, faster agent models lower the barrier to running continuous autonomous software and security agents.

The latest

Chinese developer MiniMax released M3.1 Flash on October 19, 2026, a model built for agentic workloads spanning coding, computer use and cybersecurity. The company says total cost for agent tasks runs roughly 70% below its previous generation, with flat pricing of $0.15 per million input tokens and $0.60 per million output tokens. Cached input is priced at $0.015.

Details

• The pricing: MiniMax set 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 that structure drives the roughly 70% reduction in total agent-task cost versus the prior generation, rather than any single discount.
• The speed claim: According to MiniMax, the model exceeds 200 tokens per second in everyday use, with a median of 170 tokens per second. The company attributes the gain to more aggressive sparsity in the architecture, and says online inference runs about three times faster than before.
• Agent benchmarks: MiniMax reported scores of 76.2 on BankerToolBench and 73.7% on DeepSWE v1.1, the two measures closest to tool-calling and software-engineering agent work. The company published the figures in its own blog post announcing the release.
• Security and computer use: The company reported 82% on CyberGym, a cybersecurity evaluation, and 61.7 on OSWorld V2, which measures a model operating a computer environment. MiniMax says the OSWorld V2 result ranks first among Chinese models.
• The hardest test: On Agents' Last Exam, a benchmark aimed at difficult multi-step agent tasks, MiniMax reported 29.6 and said the result leads Chinese models. The company did not publish comparative scores for non-Chinese competitors alongside its claim.
• Where to get it: Access runs through the MiniMax API, the company's M subscription plan, and the MiniMax Code desktop application. The three channels cover developers calling the model programmatically and users working inside a local coding environment.
• Open weights pending: MiniMax said the model's weights have not been released yet and will follow at a later date, without naming a timeline. A Pro version is also planned, and the company did not specify how it will differ in capability or price.
• The stealth test: MiniMax said the model users had been testing anonymously under the name Space Bunny was an early build of M3.1 Flash. The company said that build processed more than 200 trillion tokens during the blind testing period.

Between the lines

The three figures MiniMax chose to lead with — a 70% cost cut, 200-plus tokens per second, and three-times-faster online inference — all describe economics rather than raw intelligence. That framing, alongside benchmarks weighted toward tool use, coding and security rather than general reasoning, points at buyers running agents continuously, where per-token price and latency compound across millions of calls.

What's next

Watch for the open-weights release and the Pro version, neither of which has a published date, and for independent benchmark reproductions of the CyberGym and OSWorld V2 scores MiniMax reported.

Source: MiniMax blog