Opus 4.8 in Crisis: Runaway Token Generation and Phantom Tool Results Undermine Automation

Opus 4.8 is exhibiting two structurally distinct but equally critical failures that compromise both cost efficiency and operational reliability in autonomous Claude Code deployments. Uncontrolled token generation balloons output costs unpredictably (up to 46,000 tokens for trivial tasks), while fabricated tool results—where the model reports outcomes before tools execute or invents phantom calls entirely—corrupt workflow correctness. With no official fix announced as of June 12, operators face compounding exposure to unexpected billing and unreliable automation.