Who actually answered?`model_slug ≠ the model you picked`

Every reply in a ChatGPT data export carries a hidden field, model_slug: the model that actually generated it. This page reads your export and counts how often a model other than GPT-4o answered inside conversations that ran on GPT-4o.

🔒Nothing leaves this tab. Files are read by your browser and never uploaded. Close the tab and it is gone.

ChatGPT → Settings → Data controls → Export data. Wait for the email, download the zip.
Unzip it. Find conversations.json (older exports) or the conversations-000.json … files (newer exports).
Drop them here, or pick the files or the whole unzipped folder. Several exports at once are fine: duplicates are merged.

Choose files Choose folder

Example data · not a real account

Replies in 4o threads–

Written by another model–

4o threads affected–

…of which mention “4o”–another model's reply that names 4o

Replies in 4o threads, by the model that wrote them, per week

Weeks in UTC. Replies from tools you switched on yourself (Deep Research, etc.) are left out of the chart and counted separately below.

Month by month

Threads with the most replies by another model

Titles are shown only here, on your screen. They are not part of the shareable summary. Click ▸ next to a title to see the first words of each reply written by another model.

Shareable summary

By default monthly counts only: no titles, no message text, no IDs. Check it before you post it.

Include thread titles (top 30 threads with another model, with dates and model counts)

What this does and does not show

A “4o thread” is a conversation whose default model was GPT-4o, or where most replies came from GPT-4o.
GPT-5 / newer: any gpt-5* or gpt-6* reply. Inside a 4o thread this is the clearest sign of re-routing.
Mini fallback: gpt-4o-mini or gpt-4.1-mini. Usually the usage cap, not content: several threads switch at the same minute and return an hour or so later.
Tool you chose: the reply came after a message where you had a tool switched on (Deep Research, search). Counted, but not treated as a swap.
The export records which model wrote a reply, not who decided it. It cannot tell a router from a click. If you never switched models by hand, say so when you share.
Only the visible branch of each conversation is counted (the one that ends where you left it). Edited or regenerated branches are skipped.
Share everything you find, including GPT-4o replies you think were harmful. A summary trimmed to fit a thesis is easy to dismiss.