MiniMax M3.1 Flash intensifies the AI agent price war

What this is about

MiniMax released the M3.1 Flash language model on October 19, 2026. The Chinese provider is competing not only with new benchmark claims but with aggressive pricing: one million input tokens cost $0.15, one million output tokens cost $0.60, and cached input costs $0.015 according to the official price list.

This matters to developers because automated AI agents can consume far more tokens than an ordinary chat. They read files, call tools, inspect intermediate results, and repeat steps. As token prices fall, long workflows that were previously too expensive can become economical.

What M3.1 Flash actually does

M3.1 Flash is designed for software development, computer use, tool calls, and security analysis. MiniMax reports median output speed above 170 tokens per second and peaks above 200 tokens per second. It says online inference is nearly three times faster than its predecessor.

In tests published by the company, the model scores 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, and 82 percent on CyberGym. These figures come from MiniMax’s announcement. They show which tasks the provider prioritizes, but they do not replace independent and reproducible comparisons.

MiniMax also said that “Space Bunny,” previously tested anonymously on OpenRouter, was an early M3.1 Flash version and processed more than 200 trillion tokens. The model is available through the API, M Plan, and MiniMax Code. The weights are promised for later, but the official announcement gives no firm date.

Why it matters

For AI agents, the key question is not only whether a model can solve a task. Cost, speed, and the number of retries matter as well. A cheap model can be attractive to companies and independent developers even if it trails a frontier model on some tasks.

Reuters, VentureBeat, and other technology outlets covered the release on October 19. The real market shift is price pressure: anyone running agents with many tool calls now has to test whether more expensive models genuinely justify their premium through better reliability. The promised weights release also matters. It would enable local testing, independent security checks, and operation without permanent dependence on one provider.

In plain language

Think of M3.1 Flash as a fast courier service. A low price per parcel is attractive, especially when thousands are sent every day. What still matters is how many arrive correctly and how often another trip is required. Tokens are like miles traveled, while failed attempts add extra journeys.

A practical example

A developer uses an agent to sort 10,000 support cases each day. The model reads an average of 4,000 input tokens and produces 1,000 output tokens per case. That totals 40 million input tokens and 10 million output tokens a day. Under the published M3.1 Flash pricing, input would cost $6 and output another $6, or $12 before tool fees, storage, and retries.

This is only an illustrative calculation. If ten percent of cases require another pass, or if a more expensive model checks the result, the total rises. Teams should therefore compare cost per successfully completed task, not just list prices.

Scope and limits

• Most benchmark and speed figures come from MiniMax. Independent reruns under identical settings are not yet available.
• Low token prices do not guarantee low total costs. Long prompts, retries, tool charges, and quality controls can narrow the advantage.
• The model weights had not been released as of October 20, 2026. Claims about local deployment, licensing, and full auditability therefore remain provisional.

M3.1 Flash is a serious price challenge, but it is not yet a proven replacement for every existing model. The sensible approach is a limited evaluation using real tasks, fixed quality criteria, and measured cost per correct result.

SEO & GEO keywords

MiniMax, M3.1 Flash, Space Bunny, AI agents, language model, API pricing, DeepSWE, OSWorld V2, CyberGym, OpenRouter, open model weights