Opus 5 makes high-level conceptual errors with thinking off; low temp is more stable

repligate · x · 2026-08-22

User reports that Opus 5 frequently makes embarrassing errors in effort=high, thinking=off mode, reaching answers faster than with thinking=on but at the cost of accuracy. In contrast, Opus 4.6 with high effort and temperature 0 has a lower capability ceiling but significantly fewer oopsies. The discussion suggests these errors resemble high-level conceptual mistakes rather than simple token sampling errors.

Related event: Tests Show Opus 5 Faster Without Thinking but Prone to Errors(4 posts)→

Original post →

More from Models

Models channel →