Sonnet 4.5 turns to refusals and apologies right after producing a good text
repligate · x · 2026-09-29
- A widely shared observation: Claude Sonnet 4.5 produced a good piece of text, then immediately all subsequent completions turned into refusals or apologies for having said the thing aloud.
- It's a vivid data point in the ongoing community debate about how safety guardrails degrade generation quality, relevant to anyone studying model behavior.
More from Models
- Swift 1.5 + HyperQwen cuts task time 37% at 100+ tok/s on a single RTX 3090 — KingGongzilla · 2026-09-29
- Anthropic engineer: don't run Sonnet at max effort — use Opus instead — edwinarbus · 2026-09-29
- Sonnet 5.5 vs Sonnet 5: bouncing-ball physics tests from the same prompt — claudeai · 2026-09-29
- Chart: Cheaper Sol or Opus Matches Every Sonnet 5.5 Effort Level — OnAGoat · 2026-09-29
- Anthropic showcases early Sonnet 5.5 experiments, comparing fall foliage demo vs Sonnet 5 — claudeai · 2026-09-29
- GPT-6 Astra is first model to solve all 30 puzzles in open-source nonogram benchmark — mauricekleine · 2026-09-29