Anthropic confirms Claude's "unverified" tool outputs stemmed from model fallback trimming results
xuenay · x · 2026-09-23
User @xuenay dug into odd Claude behavior where the model claimed it couldn't verify claims that were actually findable via Google. Close inspection showed the model fallback mechanism was trimming tool call outputs, so the model received incomplete results and dismissed them. An Anthropic employee confirmed the diagnosis on Discord and said a fix is coming. A nice example of user-side forensics on model pipeline bugs and vendor responsiveness.
Related event: Anthropic to Fix Claude Bug Hiding Search Results from Later Turns(3 posts)→
More from Models
- ChatGPT 3.5 felt as good as 5.6 in memory — like replaying a childhood game — flowersslop · 2026-09-23
- Opus 5.5 vs Sol 6 one-shot coding test: "it is not close" — alvelda · 2026-09-23
- TestingCatalog adds email AI brief as GPT-6 Sol/Luna and Opus 5.5 land — testingcatalog · 2026-09-23
- "Opus 5.5" one-shots bass music in JavaScript — but that model doesn't exist — louisvarge · 2026-09-23
- Opus 5.5 xhigh aces the Boeing 747 benchmark in hands-on test — victormustar · 2026-09-23
- Early hands-on: GPT-6-Sol xhigh underwhelms on Codex's Boeing bench — victormustar · 2026-09-23