Anthropic warns Opus 5 may leak fake tool calls when thinking is off
heypearlai · x · 2026-07-29
Opus 5 can silently break agent loops without thinking
The thread says that with thinking disabled, Opus 5 may emit a tool call as plain text instead of a real call. In that case nothing executes, the loop takes no action, and the fake call remains in history as if it had happened.
It also warns that the same mode can leak internal tags into visible output. According to Anthropic, telling the model never to output those tags can make the leak worse.
Recommended fix:
- keep thinking enabled
- lower the effort level instead of disabling thinking
The post also notes that Opus 5 can narrate its own corrections mid-response, so structured outputs should explicitly ask for only the final answer.
Related event: Prompting Tips and Pitfalls for Claude Opus 5(4 posts)→
More from coding & agent
- 'LLM code is sloppy' rebutted with real ffmpeg code: identical if/else branches in a top library — kingslyj · 2026-10-16
- Agent actions need checks outside the model: NVIDIA demos lean on guardrails amid wave of real agent incidents — Creamy-And-Crowded · 2026-10-16
- Vitalik builds "paranoid mode" agent wrapper with Tor, zkAPI and local prompt rewriting — DavideCrapis · 2026-10-16
- Agent fine-tunes Qwen3.5-0.8B with synthetic data: 4.7%→76.3% accuracy, 11.9× faster — ben_burtenshaw · 2026-10-16
- Beam-501B agent demo: solo beats 5-agent team on ProgramBench rebuild task — ben_burtenshaw · 2026-10-16
- Beam agent builds BeamSpeak, a local Parakeet dictation app for macOS — ben_burtenshaw · 2026-10-16