Watch what people keep deleting
Repeated edits are evidence about the default model.
Retrospective field note for 2026-10-05, reconstructed on 16 November 2026. The archive date is not a claim of publication at the time.
If every answer loses its first paragraph during editing, the first paragraph is probably a product problem.
Track those repairs. Unsupported claims removed. Constraints restored. Openings shortened. A draft can be accurate and grammatical while repeatedly missing the audience or the required shape.
Use those patterns in evaluation. Sanitise examples while preserving the difficult structure: an ambiguous reference, a missing field, conflicting requirements, a request that only needs one sentence. Do not copy private user material indiscriminately into a test collection.
The default model handles work when nobody wants to think about model selection. It should cope with ordinary requests without requiring an elaborate prompt every time. It should also recognise when the evidence is insufficient instead of padding the answer around the gap.
Sonnet 5.5 is another candidate. Compare accepted outputs and the edits needed to reach them. The launch demo may show the ceiling; routine repairs show what using the model every day is likely to cost.
Anthropic introduced Claude Sonnet 5.5 on 28 September. Source.