MiniMax M3.1 Flash costs half of M3 per token, and its weights are promised

MiniMax has added M3.1 Flash to the MiniMax API release notes, calling it an ultra-efficient open-weight model for everyday work. It costs $0.15 per million input tokens and $0.60 per million output tokens, and the company says weights will follow later.

The entry appeared on the release-notes page on October 19, 2026, as a short line linking to a MiniMax blog post. The post carries the specifics.

Flat pricing and three ways to get it

According to MiniMax, M3.1 Flash is available now through the MiniMax API, the M Plan and MiniMax Code. Pricing is flat per million tokens: $0.15 for input, $0.60 for output and $0.015 for cached input. The company also says total cost for agentic tasks is 70% lower than with its predecessor.

On speed, MiniMax claims inference is nearly three times faster than before. It cites output of more than 200 tokens per second in everyday use, with a median of more than 170.

The model is described as open-weight, but the weights are not out. The blog says only that they “will be released later.” It gives no date and does not state a license. It also does not publish a parameter count or a context window. Anyone who wants to self-host, fine-tune or run the model in their own region has to wait. That matters because the label “open-weight” currently describes an intention, not a download.

Benchmarks MiniMax chose to show

The blog lists these results for M3.1 Flash:

• BankerToolBench: 76.2
• DeepSWE v1.1: 73.7%
• OSWorld V2: 61.7, which MiniMax says is first among Chinese models
• Agents’ Last Exam: 29.6, also first among Chinese models
• CyberGym: 82%

On AutomationBench-AA, MiniMax says only that the model is among the top performers, with no score. These are the company’s own figures. The page names no comparison models for most tests, and the claims have not been independently verified. The “first among Chinese models” ranking is also limited by the fact that the comparison set is not listed.

The blog frames the release modestly: “Flash represents our ongoing exploration of the balance between efficiency and capability.”

Half the price of M3, level with the cheapest tiers

The analysis here is about numbers in context. Our AI API pricing tracker, which uses OpenRouter listings, shows MiniMax M3 at $0.30 input and $1.20 output per million tokens, with a 1 million token context. M3.1 Flash is therefore priced at exactly half of that on both sides. M3 launched on June 1, 2026, according to the same release-notes page, so the Flash tier arrives under five months later.

Against other small models the picture is less dramatic. The same tracker lists OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 at $0.10 input and $0.50 output. Those are cheaper than MiniMax’s new price. It also lists DeepSeek V4.1 Flash 1017 at $0.15 and $0.60, the same price as M3.1 Flash, and Alibaba’s Qwen3.8 Flash at $0.15 and $0.47. So M3.1 Flash is not the cheapest option. Its pitch rests on whether agent quality at this price holds up, which only outside testing can settle.

The cached-input rate is the more interesting figure for agent work. At $0.015, repeated context costs a tenth of the normal input price. Coding and computer-use agents resend long histories on every step, so that rate drives much of the real bill. MiniMax’s 70% claim presumably reflects this, but the blog does not show how it was measured or against which workload.

Regional access is another practical point. OpenRouter’s documentation for the model says it is not available through EU in-region routing, which may matter to European teams with data residency rules. That comes from OpenRouter’s page, not from MiniMax.

What is still missing

Several things are not established by the sources read. There is no weights date, no license, no architecture or size, and no context length. The 70% cost reduction and the threefold speed gain lack named baselines. For earlier MiniMax work on a different product line, see our coverage of Speech 2.5.

For now, developers can call the model through the API and judge it on their own agent workloads. The open-weight question will be answered when MiniMax publishes the files and the terms that go with them.