GitHub Copilot context window size difference of over 300K when switching models.

This started out as a bug report, but now it's more of a puzzling question.

1. Disable summarization for a workspace.
```json
…
"settings": {
"github.copilot.chat.summarizeAgentConversationHistory.enabled": false
}
…
```
2. Work all day with Claude Opus 4.8 with a 1M context window, until the context window is about halfway filled up. (I didn't note the exact number.)
3. Switch to GPT-5.6 Luna, also with a 1M context window, and issue a prompt.

**Expected:** The context window of GPT-5.6 Luna remains about half full (well over 400K).

**Actual:** The context window of GPT-5.6 Luna is only about 140K.

4. Interestingly, I switched back to Claude Opus 4.8, the context window again showed over 500K!

So maybe nothing was summarized or dropped. But what explains the discrepancy when I switch to another model? Does GPT-5.6 Luna truly have that much of a more efficient tokenization to cause a 300K difference in context window size?