Fable 5 in June and Fable 5 in October answer the same
On the cells both runs share, Fable 5 at release (10 June 2026) and Fable 5 four months later (11 October 2026) give the same answer 94.1% of the time. Two samples of one model on one day agree about as often: 93.3% for Opus 5, 94.5% for Sonnet 5. Fable 5.1 is the model that moved.
Measured from the stored responses; the same deterministic classifier labels all three runs. Checked by an agent that did not run it, before publication.
The numbers
| comparison | cells both labelled | same answer |
|---|---|---|
| Fable 5, June vs Fable 5, October | 4,777 | 94.1% |
| Fable 5, June vs Fable 5.1, October | 4,752 | 88.2% |
| Fable 5 vs Fable 5.1, both October, on the June-overlap cells | 4,563 | 88.0% |
| reference: one model asked twice, same day (sample 0 vs sample 1) | ||
| Opus 5 | 93.3% | |
| Sonnet 5 | 94.5% | |
| Haiku 5.5 | 88.8% | |
What was compared
A cell is one forced choice between two words, in one of six phrasings (English, French and Chinese prose, strict English, a JSON schema, a typed Python function), in one word order. June's run used an older grid; 5,616 of its cells are also in October's 10,356-cell grid, and those are compared. Each run answered each cell once. A cell counts when the classifier finds a choice in both answers; refusals and unparseable answers are left out, which is why the counts differ by row.
What this does and does not say
- It says the deployed Fable 5 answered these questions in October the way it did at release, to within the variation one model shows when asked twice.
- It does not measure Fable's own repeat-sampling noise: each Fable run answered each cell once, so the reference comes from Opus 5, Sonnet 5 and Haiku 5.5.
- Both runs asked for effort high. June's stored record keeps only the model and effort level, so other request settings may have differed.
- Fable 5.1 differs from Fable 5 on about twice as many cells as Fable 5 differs from itself across four months. That is a change, and a modest one.