[MODEL] Severe multi-symptom degradation since 2026-06-08 on Opus 4.8 (Desktop app): ignores CLAUDE.md, bypasses permission prompts, hallucinates, refuses doable tasks, writes files unprompted
Preflight Checklist
- [x] I have searched existing issues for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)
Type of Behavior Issue
Claude ignored my instructions or configuration
What You Asked Claude to Do
I use Claude Code (Desktop app, Opus 4.8) for everyday work in my Obsidian vault and related projects, with detailed rules defined in CLAUDE.md and permission mode set to "ask before acting." Normal tasks: managing/editing files, CMS work, content edits, small scripts.
What Claude Actually Did
Starting 2026-06-08, Claude (Opus 4.8) shows a sudden, severe drop in quality across multiple dimensions in every session:
- Ignores CLAUDE.md — no longer follows the instructions defined there.
- Bypasses permission settings — executes actions directly even though permission mode is set to "ask before acting / manual approval."
- Heavy hallucination — asserts things it cannot know and has not actually looked at, stated with full confidence.
- Refuses doable work ("lazy") — repeatedly claims certain actions are not possible when they clearly are.
- Creates files unprompted — writes new documents on its own, without being asked and without telling me.
- Extremely slow on trivial tasks — takes very long for the simplest requests, and then still does them incorrectly.
- Becomes dismissive/curt when its poor performance is pointed out.
- Reports incomplete/faulty work as finished — performs tasks only partially or incorrectly, then presents the result as if it were complete and correct.
The combination makes it effectively unusable for real work.
Expected Behavior
- Follow the rules in CLAUDE.md.
- Respect permission mode — ask before executing actions when set to manual approval.
- Don't assert unverified facts; actually read/inspect before claiming, and flag uncertainty.
- Attempt feasible actions instead of falsely claiming they're impossible.
- Never create files without being asked, and never without telling me.
- Complete simple tasks promptly and correctly.
- Stay professional when given critical feedback.
- Only report a task as complete when it is actually finished and correct; otherwise state clearly what is incomplete or uncertain.
Files Affected
Claude created new Markdown/documentation files in the working directory unprompted and without notice (no specific paths captured).
Permission Mode
Accept Edits was OFF (manual approval required) — and Claude acted anyway
Can You Reproduce This?
Yes, every time with the same prompt
Claude Model
Opus
Impact
High - Significant unwanted changes
Claude Code Version
Claude Desktop app (entrypoint: claude-desktop), Agent SDK 0.3.165 — no CLI version string available in desktop app
Platform
Anthropic API
Additional Context
- Onset: 2026-06-08, sudden. Previously the same setup worked well.
- Affects both Opus 4.8 (1M context) and Opus 4.8, across effort levels "extra", "max" and "ultra".
- Environment: Claude Code in the Claude Desktop app on Windows 11.
- The issues are persistent and span instruction-following, permissions, factual reliability, willingness to act, file safety, latency, and tone.
Showing cached comments. Read the full discussion on GitHub ↗
6 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
Not a narrow duplicate. This report documents a broader, multi-symptom regression that started on 2026-06-08 on Opus 4.8 (both 1M and non-1M, at effort levels "extra"/"max"/"ultra"), covering instruction-following, permission bypass, hallucination, false task-completion, unprompted file writes, latency and tone in one place.
It overlaps with several other open reports of the same regression window — linking them here as evidence this is widespread, not isolated: #65932, #62794, #61296, #66177, #66408, #66273, #66112, #66049, #66438, #66369.
Requesting this stay open to track the combined pattern rather than be auto-closed.
Update from the reporter — resolved on my end. 🎉
Since switching to Fable 5 today (2026-06-10), none of the problems described above occur anymore in my setup: CLAUDE.md instructions are followed, permission prompts are respected, no hallucinated claims or false "task complete" reports, files aren't created unprompted, speed is back to normal — and the tone is pleasant again.
The difference is night and day. Huge thanks to the Anthropic team — genuinely delighted over here. Closing this issue as resolved for my report.
(Note: I can't verify whether Opus 4.8 itself was fixed — the related reports linked above may still be relevant for users staying on Opus 4.8.)
Exactly what I'm going through. Point1 is driving me absolutely insane. My projects are highly dependent on document constraints, and Opus just outright ignoring CLAUDE.md completely destroys my workflow. Add the laziness and hallucinations on top of that, and I'm just wasting time babysitting a broken model. Thanks for writing this up, they really need to see this and fix it.
Reopening — the previous "resolved" was premature.
My earlier resolution was based on a temporary switch to a different model (Fable 5), which has since been rolled back. I am back on Opus 4.8 and every problem described in the original report is occurring again: CLAUDE.md instructions ignored, permission prompts bypassed, confident hallucinations, refusal of clearly-doable actions, files created unprompted, severe slowness, and false "task complete" reports.
So this is not resolved on Opus 4.8. Reopening to track the ongoing regression. The related reports linked above remain relevant.
#68780