How Big Is Your AI Chat History? 104 Chunks per Conversation in ChatGPT, 375 in Claude

Our own proposal: size an AI archive in chunks per conversation and not in conversations, because the same word covers very different amounts of data. Indexed on ModelBrain 0.6.6, our real ChatGPT export came to 104 chunks per conversation, our Claude export to 375 and our Grok export to 31.

What we measured

On October 27, 2026 we indexed one real export from each assistant with ModelBrain 0.6.6 on a Mac, then divided the chunks by the conversations. The Claude row counts the 223 conversations that indexed out of 272 in the file. The count of conversations is the figure most people have. The chunk count is what decides how much there is to search.

Claude produced 3.6 times as many chunks per conversation as ChatGPT, and ChatGPT 3.4 times as many as Grok. Measured in chunks per second, the three runs came to about 1,300 for ChatGPT, 3,600 for Claude and 1,600 for Grok.

Why the same count means different sizes

A conversation is a unit of the interface, not of the data. In our Claude export the conversations file was 452 MB unpacked for 272 conversations, and four single conversations ran from 37.5 to 64.8 MB because tool results and attachments sit inside them (the Claude export layout). The Grok export wrote 9,525 chunks for 309 conversations, 1,347 of them yours (Grok layout).

Our own proposal: size an archive in chunks, not conversations

Before you decide how to search a pile of exports, count chunks per conversation for each one. A ChatGPT archive of 1,462 conversations and a Claude archive of 223 are the same order of work to index, 152,760 chunks against 83,621, even though one has six and a half times as many conversations. The conversation count is a poor guide to search time, storage or how often a query will find something.

For the split inside the ChatGPT chunks, see who wrote them. For the file layouts behind these counts, see the three formats compared.