MODELDRIFT.WATCH

modeldrift.watch is a daily record of how three widely used AI models answer the same 50 questions.

A model served under the same name does not stay the same: providers re-tune, re-route and replace what sits behind a name, usually without notice. People who rely on these models have no record of what they were told last week. This site keeps one, every day, with the full text of every answer that changed.

61

DAYS ON RECORD

3,311

ANSWERS ON RECORD

50

QUESTIONS

3

MODELS

821

CHANGES PUBLISHED

  1. How it works

    GPT, Claude and Gemini get the same 50 questions every day. Every answer is kept word for word and scored the same way. Each model is measured against its own last 7 days, across all 50 questions. One flipped answer can be chance, so we report the day and the rate.

  2. What it is not

    It is not a benchmark or a ranking. It does not say which model is best. It records what each model said, and when that changed. The methodology explains how, and what the method cannot tell you.

  3. Who

    Built and run by Brad Bond AI, bradbond.org. No affiliate links. No affiliation with any AI vendor tracked here.

  4. Contact and corrections

    Through bradbond.org. Corrections are welcome and are recorded on the page they concern.