MiniMax says M3 sent a business email for $0.018

MiniMax, led by founder Yan Junjie, said M3 created an inbox, wrote a business email and sent it for $0.018 in an August 25th post on X. MiniMax called M3 the cheapest model in a four-model demonstration and said it completed the assignment.

MiniMax on X

The claim packages Yan's long-running argument about AI economics into an unusually legible number: 1.8 cents for a finished piece of work. That framing gives MiniMax an opening in a market where buyers increasingly compare model capability with the cost of long tool-use loops.

Yan has spent his career on the technical side of that tradeoff. Before founding MiniMax, he worked at SenseTime from 2015 until 2021, progressing from intern to vice president and senior research roles. He studied mathematics at Southeast University, earned a doctorate in artificial intelligence from the Institute of Automation at the Chinese Academy of Sciences and conducted postdoctoral research at Tsinghua University. Yan now holds all four of MiniMax's central jobs: chairman, CEO, CTO and founder. (MiniMax management page)

The email demonstration still falls short of a benchmark that buyers could reproduce. The attached video presents four terminal-style Hermes Agent panes labeled Claude Opus 5, MiniMax-M3, GPT-5.6 and Gemini 3.7 Flash. An Atomic Mail banner identifies the inbox service, and the Hermes interface shows its atomicmail skill as active.

The visible frames do not show the final $0.018 figure. Each pane displays a $0.000 counter, while the post supplies the claimed completed-task cost. The terminal also warns that the Hermes installation is 2,562 commits behind and recommends an update. That does not invalidate the run, but it makes the exact software environment important to any attempted replication.

The metric Yan wants the market to use

MiniMax released M3 on June 1st, nearly three months before the email post. MiniMax describes M3 as a natively multimodal coding and agent model with a context window of up to one million tokens. Its published weights list roughly 428 billion total parameters and 23 billion active parameters, with text, image and video input support. (MiniMax's M3 announcement)

The model's architecture supports Yan's commercial pitch. Long-running agents accumulate prompts, tool results, files and intermediate reasoning, turning token consumption into an operating expense. A model that completes an entire task cheaply can be commercially useful even when a rival posts a higher score on a conventional reasoning test.

Yan outlined that thesis in a 2025 conference keynote. He argued that falling per-token costs would be accompanied by sharply rising token consumption as agent interactions expanded from short conversations into complex assignments. MiniMax's $0.018 email is a compact demonstration of that idea: price the finished workflow, then ask how many workflows a customer can afford to run. (MiniMax's WAIC keynote)

That is a useful direction for agent evaluation. API price tables reveal the cost of inputs and outputs. They do not capture retries, failed tool calls, unnecessary browsing, human review or a model abandoning the task before delivery. A cheap model that needs five attempts can cost more per successful job than an expensive model that finishes on the first try.

A practical demo with benchmark-sized gaps

A repeatable completed-work benchmark needs the exact prompt, model settings, provider route, Hermes Agent version, mailbox pricing, token usage, wall-clock time and success criteria. It also needs repeated runs. One successful completion cannot show how often M3 fails, whether another model produced a better email, or how retries change the average cost.

The accounting boundary matters just as much. MiniMax's post does not specify whether $0.018 covers only model inference or also the inbox, tool calls and surrounding infrastructure. The Atomic Mail branding makes the task concrete, while leaving the reader without a full bill of materials.

The email post offers none of the methodological detail needed to reproduce the result. (MiniMax's M3 announcement)

MiniMax would strengthen its argument by releasing the run as a small test package: pinned software commits, task instructions, cost logs and multiple trials for each model. The task itself is well chosen. Provisioning an inbox and sending a usable message is easy to understand, requires real tool use and ends with an external artifact that can be checked.

MiniMax turns efficiency into distribution

RuntimeWire reported on August 21st that MiniMax had teased M3 on SambaNova without publishing the model setup or a result. On August 24th, MiniMax put M3 and three other models on a two-week GMI Cloud trial, again using outside infrastructure to widen access.

The Hermes Agent video advances that strategy with a result people can remember. Few prospective customers can interpret a sparse-attention diagram or translate a coding benchmark into an automation budget. A completed business email for $0.018 takes no conversion.

That is why the benchmark details matter. MiniMax benefits when model competition moves toward the cost of completed work, and Yan has built M3 around the premise that intelligence must become cheaper to reach a larger market. The email run is consistent with that bet. A reproducible test would show whether the number belongs in an operating plan rather than a promotional post.