[BUG] Fable 5 quality regression: ignores explicit instructions after acknowledging them, increasing hallucinations

Status Open
Reported on v2.1.235
Maintainer reply None cached
Activity 0 comments · opened Aug 19, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

Environment

  • Model: claude-fable-5 (Effort: High)
  • Clients affected: BOTH the Claude Code VS Code extension (anthropic.claude-code, v2.1.235) AND the Claude desktop app (Code feature). Same degraded behavior reproduced in both, so this is model-side, not client-specific.
  • Date observed: August 2026, worsening over recent weeks

Summary

Fable 5's instruction-following and factual reliability have degraded noticeably. The model acknowledges explicit instructions, claims to have committed them to memory/rules, and then violates them within the same session. Hallucination frequency has also increased.

Additional symptoms

  • Hallucinated facts stated with confidence (e.g., wrong institutional counts/details in document review tasks), getting more frequent
  • Instructions given early in a session are silently dropped as the session grows, even after explicit acknowledgment
  • Over-interpretation of simple constraints into broader restrictions the user never asked for

Impact

This makes Fable 5 unreliable for multi-step document workflows. The acknowledge-then-violate loop is worse than a refusal, because it wastes turns and destroys trust in the model's confirmations.

(Screenshots attached below)

What Should Happen?

An acknowledged instruction should be followed for the remainder of the session. A simple constraint ("don't auto-open files") should not be expanded into unrelated restrictions the user never asked for.

Error Messages/Logs

Steps to Reproduce

  1. In a document-review session (VS Code extension or desktop app Code feature), explicitly instruct the model NOT to open files on my behalf when delivering outputs - just list the absolute paths.
  2. The model acknowledges, creates a rule file (feedback-no-relative-chat-links.md), and confirms the rule was saved.
  3. Continue the session. The model then over-corrects in the wrong direction (strips path usability entirely) while still not fixing the actual behavior complained about.
  4. Instead of changing behavior, it produces a performative artifact (a new .md rule file) - a pattern where the model "evaluates" instructions rather than executing them.
  5. Repeat corrections -> same acknowledge-then-violate loop.

Claude Model

Other

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.235 (VS Code extension); also reproduced in Claude desktop app

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_

View original on GitHub ↗