davidad: Opus 4.6 is the most capable Claude supporting temp=0
davidad · x · 2026-08-22
Safety researcher davidad compared Claude Opus 5 vs Opus 4.6: running Opus 5 with effort=high and thinking=off reaches right answers faster than with thinking on, but "constantly oops-ing all over the place." Opus 4.6 with the same settings plus temp=0 has a lower capability ceiling but far fewer blunders, making it his pick for some use cases.
Related event: Tests Show Opus 5 Faster Without Thinking but Prone to Errors(4 posts)→
More from Models
- Anthropic opens Mythos 5 access for Claude Enterprise security beta — AccBalanced · 2026-08-22
- Rumor suggests Ox Alpha stealth is Cursor Composer based on GLM 5.2 — teortaxesTex · 2026-08-22
- OpenAI Launches Ultrafast GPT-5.6 Sol with 14x Speed Boost — gajesh · 2026-08-22
- Users Report Degradation in OpenAI Sol High Chat/Codex — Illustrious-Bet-1368 · 2026-08-22
- Discussion: Which Public LLM Benchmarks Do You Actually Trust? — ThomasAger · 2026-08-22
- Ornith 1.5 32B tested: Fast performance in Three.js demo — draginol · 2026-08-22