BUG PLEXOR TEST: Opus escalation generates minimal output for high cost

Resolved ๐Ÿ’ฌ 3 comments Opened Jan 9, 2026 by micoverde Closed Jan 9, 2026

Bug Bash - Plexor Routing Analysis

Severity: HIGH ๐Ÿ”ด

Description:
During live monitoring of Plexor staging, observed that Claude Opus 4.5 was invoked for simple continuation tasks, generating only 9 output tokens while consuming ~28K input tokens.

Evidence from logs:

[ANTHROPIC REQUEST] model=claude-opus-4-5 max_tokens=16000
[ANTHROPIC RESPONSE] tokens: prompt=27909 completion=9 total=27918
[ANTHROPIC RESPONSE] cost=$0.139770
[ANTHROPIC RESPONSE] response: continue fixing the remaining input fields

Impact:

  • $0.14 per request for 9 output tokens ("continue fixing the remaining input fields")
  • Two consecutive Opus requests = $0.28 for 18 output tokens
  • Haiku was handling the same task for ~$0.028

Root Cause Investigation Needed:

  1. Why did routing escalate from Haiku to Opus mid-task?
  2. Was this user-requested (passthrough mode) or automatic?
  3. Should minimal-output responses trigger cost optimization?

Recommendation:

  • Add guardrail: If completion tokens < 50 and cost > $0.05, flag for review
  • Consider output-aware routing that can downgrade mid-stream

---
Found during real-time Plexor forensic analysis session.

@claude and @codex please review this pr

View original on GitHub โ†—

This issue has 3 comments on GitHub. Read the full discussion on GitHub โ†—