A bug in Anthropic's Claude Opus 4.8 model can silently corrupt tool calls, particularly in Japanese environments and during long-running sessions. This corruption causes function arguments to appear as raw text and function tags to become nonsensical tokens, creating an self-reinforcing loop where the model imitates the corrupted output. The issue, which does not affect Opus 4.7, can be mitigated by downgrading to the previous version, decomposing tasks to keep sessions shorter, or using the `/compact` command to summarize history. AI
IMPACT This bug highlights potential risks in long-context models and the importance of task decomposition for reliable AI operation.
RANK_REASON A user-reported bug in a specific model version with detailed technical analysis and mitigation steps. [lever_c_demoted from research: ic=1 ai=1.0]
Read on dev.to — Claude Code tag →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →