Haiku 5.5 nails every checklist, drops the point

The test was set up to isolate the model itself, not the retrieval pipeline. Fourteen web articles were fetched and saved on 10/9, and the identical files were fed to every condition, so differences could not come from CAPTCHA or timeouts changing what each model read.

A line is drawn between two kinds of summary failure. A factual error can be caught by checking what is written against the source. A missing core claim cannot be caught that way, because nothing about it appears in the summary for a checker to compare. The plain-prompt Haiku summaries passed the factual check at full marks in the top row of the comparison table while still dropping the central point in 4 of 13 cases — the failure sits outside the net.

Retrieval limits behaved differently. Opus 5.5, left with the plain prompt, wrote about missing images, truncated bodies, and CAPTCHAs in only 3–4 out of 10 articles, showing that gap is not unique to cheaper models. The prompt step closed both gaps at once, pushing Haiku's core-claim score to 13/13 and both models' retrieval-limit score to 9–10/10. An example was the GPT-6 announcement, where the plain Haiku summary listed every product detail correctly but never mentioned the closing line about software adapting to people; the stepped version put that line at the top and flagged that the interactive UI examples in the fetched text were labeled GPT-5.6 Instant.