[Bug] Anthropic API Error: False-positive Usage Policy block on biomedical research with Opus 4.8

Resolved 💬 4 comments Opened Jun 5, 2026 by AkiraLan Closed Jul 12, 2026

Bug Description
Title: False-positive Usage Policy (AUP) block on legitimate biomedical research â^@^T Opus 4.8 only; Opus 4.7 unaffected

Severity: High â^@^T blocks a legitimate academic research workflow

Environment

  • Claude Code (native install), versions tested: 2.1.161 / 2.1.163 / 2.1.165
  • Model where it occurs: claude-opus-4-8[1m] (started after the Opus 4.8 migration on 2026-06-05)
  • Model where it does NOT occur: claude-opus-4-7[1m]
  • Context: single-cell RNA-seq study of mouse atherosclerosis drug-target discovery; a memory plugin (claude-mem) injects prior project notes into the system context each session.

Summary
Since migrating to Opus 4.8, requests intermittently fail with:
API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy ...
The content is entirely legitimate academic biomedical research (regulated cell-death pathways, druggable-target nomination, treatment-response analysis in cardiovascular disease). Switching the model to Opus 4.7 resolves it completely with identical context.

Request IDs

  • req_011Cbk5a5qFg2yQdVNDHvbR1
  • (additional blocked request IDs available in my error history on request)

Isolation performed (what was ruled OUT as the cause)

  1. Specific memory records ^T deleted injected notes day-by-day (incl. the densest); still triggered. Not a single record.
  2. Injection volume/density reduced injected items from 50 to 8; still triggered on 4.8.
  3. Third-party plugins (oh-my-claudecode) fully disabled; still triggered.
  4. Claude Code client version downgrading (2.1.165 ^R 2.1.163 2.1.161) only delayed the trigger (instant after several turns), never eliminated it on 4.8.
  5. Confirmed necessary condition: with the memory injection OFF the session is always fine; with it ON it fails on 4.8 but never on 4.7.

Conclusion
The trigger is model-coupled: Opus 4.8's input safety classifier appears to false-positive on dense legitimate biomedical terminology (cell-death programs + drug/target + treatment-reversal language) when accumulated in context. Opus 4.7 does not. Client version only changes the headroom, not the root cause.

Impact
Legitimate, authorized academic pharmacology / single-cell genomics research is being blocked on Opus 4.8. The workaround (pin Opus 4.7) sacrifices the newest model.

Requested action
Please review the Opus 4.8 input-content classifier for false positives on biomedical research contexts (regulated cell death, druggability, disease-reversal pharmacology). Happy to provide full reproduction context and additional request IDs.

Environment Info

  • Platform: linux
  • Terminal: tmux
  • Version: 2.1.163
  • Feedback ID: c9fe0137-30f0-4497-9588-69c9395b19f7

Errors

[]

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗