Independent test finds MiniMax M3.1 Flash cheaper but not faster than M3

A vendor's claim of a faster, cheaper flash model did not survive one small test on real agent work. Jessica Wachtel at The New Stack ran MiniMax M3.1 Flash and M3 on three agent tasks and reports the result. Flash scored the same as M3 on the two graded tasks. The estimated total cost for Flash was about 44 percent lower across the three tests. Flash's total wall time was 661.6 seconds. M3 took 465.0 seconds. Neither model answered the third task, a long-context question.

What MiniMax says about Flash

MiniMax released M3.1 Flash on October 19, 2026. The company says Flash cuts the overall cost of agent tasks by 70 percent. It says online inference runs nearly three times faster than M3. MiniMax lists Flash at $0.15 per million input tokens and $0.60 per million output. Those are half M3's discounted rates for the same request size. Artificial Analysis independently scored Flash at 72.5 percent on AutomationBench-AA, a multi-step business app test.

What the three tests measured

Wachtel ran both models once per task through OpenRouter, pinned to MiniMax's own servers. Each run used default settings. Task one was an agentic bug fix on a Python job-queue library with three interacting bugs. Task two was an incident investigation across logs and 20 fake services. Task three was a long-context test on a 500,000-token wiki.

Both models scored 5 of 6 on the bug fix and 10 of 10 on the incident. Flash used more than twice the input tokens on the bug fix. Its estimated cost on that task was about 23 percent lower than M3. On the incident test, Flash's cost was about 68 percent lower, close to MiniMax's 70 percent claim. Both models failed the long-context task on the first run and on a retry with lower reasoning.

What this tells a buyer

The ratio MiniMax quotes is cost per task, not time to task. Flash was cheaper by estimated price in both passing tests. Flash was slower by wall time on the whole run. A team paying for the model gets a lower invoice. A team waiting for an agent gets a longer answer. Flash's tool style explains part of the gap. It read 37 files over its first 14 turns on the bug fix before making an edit. M3 ran the tests once and started fixing on turn 4.

Wachtel used OpenRouter's free Flash endpoint, so her Flash prices are MiniMax list prices, not what she paid. She says three one-off tests cannot show whether the vendor's benchmark claim holds. A buyer should run the same agent on both models before cutting a cheque.

Source

The New Stack.