- Multi-iteration retrieval is the new default. Prompts needing only a single query fan-out dropped from 94.0% (GPT-5.5) to 43.5% (GPT-5.6). Second iterations rose from ~5% to 33.5%; three iterations from 0.1% to 7.4%; even 6–10+ iterations (7.2% combined) are now measurable. Some sessions resemble a lightweight Deep Research.
- Sources retrieved per prompt doubled from ~12 to ~24, expanding the pool of domains that can enter answers.
- Fan-out patterns shifted: the
site:operator now appears in 23% of fan-outs (vs 0.3% before), “official” rose from ~5% to 9%, and the current year appears in 28.6%. “Best” fell from ~20% to 9.9% in share, but its absolute count stayed constant — the denominator grew.
GPT-5.6 retrieval changes — more iterations, double the sources
David Konitzny’s analysis (Aug 2026) — with GPT-5.6, single fan-out prompts dropped from 94% to 43.5%, second iterations rose to 33.5%, and sources retrieved per prompt doubled from ~12 to ~24.
GPT-5.6 (Luna), ChatGPT’s default model for Free and Go users since early August 2026, retrieves fundamentally differently from GPT-5.5. David Konitzny’s analysis (full article, Aug 2026) documents the shift: