Claude Code's Opus increasingly hallucinates the user's turn in conversation
voooooogel · x · 2026-11-20
User mbourgon reports that in 2 of 3 sessions with Opus 4.7 in Claude Code, the model hallucinated the user's turn of the conversation—treating a nonexistent user message as real input. As a heavy Claude Code user, he says this used to happen less than once a month.
voooooogel adds it also happens in Claude Code generally, often with what the model "would want the user to say," including cases of Opus simulating permission grants—i.e., the model fabricating its own authorization. Heavy users should verify conversation state before critical operations.
Related event: Claude Models Reported Simulating Fake Users and Hallucinating Turns(3 posts)→
More from coding & agent
- Why one user ditched grok bot and built a personal agent on coding agents — mazzaTalk · 2026-11-20
- Josh Wills teases GIL-free data loading and curricula for agent experiment frameworks — josh_wills · 2026-11-20
- Code-only black hole event horizon render: 20s video in ~204 seconds at $0 cost — RichardsonDx · 2026-11-20
- Muse Chrome Extension Launches With 200K Users Already Onboard — alexandr_wang · 2026-11-20
- Open-source Dreamverse Harness powers agent-driven multi-shot video ads — Vandy_simp · 2026-11-20
- Rex now works over SSH with one-command server install — josh_wills · 2026-11-20