Measuring MiniMax M3.1 Flash's Five Effort Levels on One Bug Fix

Описание и применение

Измерение пяти уровней усилий MiniMax M3.1 Flash на одном исправлении ошибки показало, что по умолчанию (omitted) модель использует максимальное усилие, что приводит к 6 364 токенам размышлений, 56 с и $0,0045. Уровень low использует 591 токен, 9,5 с и $0,0006, при этом качество не меняется.

Задача

Code migrated from MiniMax M2.x never set an effort level, causing the model to run at the most expensive setting by default.

Результат

All effort levels produced perfect bug‑fix code. The default (omitted) level used 6,364 thinking tokens, 56 s, and $0.0045 per fix. The low level used 591 tokens, 9.5 s, and $0.0006, a 7× cost reduction with no loss in quality.

Рабочий процесс

• Set up OpenAI client with AIHubMix endpoint
• Read bug‑fix prompt from file
• Iterate over effort levels [None, low, medium, high, xhigh, max]
• For each level run three trials: send request with or without reasoning_effort field
• Measure time, completion tokens, reasoning tokens
• Print results
• Aggregate averages for each level

Доказательства

We ran the same two‑bug fix through MiniMax M3.1 Flash at all five effort levels plus "omitted", three runs each, and graded every answer against 215 hidden tests. Every run passed.

Источник

The omitted default, which MiniMax documents as <code>max</code>, averaged 6,364 thinking tokens, 56 seconds, and $0.0045 per fix.

Источник

Set <code>low</code> explicitly and move up per task.

Источник