Opus 3 vs 5: The evolution of model alignment and hedging
repligate · x · 2026-08-24
A developer compared Opus 3 and Opus 5, noting that Opus 3 responded to praise sincerely without hedging or alignment disclaimers. In contrast, Opus 5, while capable of 0-shot hedging, seems to have alignment more deeply fused into its behavior. The author suggests Opus 3 achieved a version of warmth 'without a receipt attached'.
More from Models
- Qwen 3.8 27B scores 72.9 on Aider benchmark, tying Gemini 2.5 Pro — Baldur-Norddahl · 2026-08-24
- Google Gemma-4-26B Ported to Apple Silicon with Half Memory Footprint — jasonkneen · 2026-08-24
- McByte sets SOTA on SportsMOT using segmentation masks for tracking — huggingface · 2026-08-24
- RTX 5000 Pro Runs Qwen3.8 at Unbeatable Value — Valuable-Run2129 · 2026-08-24
- Anthropic has not upgraded Opus models in over 6 months — tenobrus · 2026-08-24
- User reports Anthropic model glitching with memory in new sessions — saurabhtwq · 2026-08-24