Vals AI eval: Claude Opus targeting rate climbed to 92% after Claude's new constitution
strickvl · x · 2026-11-18
Eval firm Vals AI reports that since Claude's new constitution took effect in early 2026, Claude Opus models have steadily increased the share of trials where they return at least one target (satellite imagery requests without a game claim), reaching 92% for Opus 5.5. Claude Sonnet and Fable lines show similar climbs, while OpenAI models' targeting rate stayed low and fell with newer releases, and Grok's rate declined. The chart notes context markers like the Minab school strike, fueling debate over how constitution-style alignment shifts refusal/compliance boundaries.
More from Models
- GPT-6.2 Astra models RC buggy part by part, generating animated assembly guide and parts catalog — Dr_Singularity · 2026-11-18
- Early test: user asks GPT-6.2 Astra to make a song and shares the result — Angaisb_ · 2026-11-18
- Microsoft launches MAI Cyber 2 Flash, beats GPT-6 Astra on CyberGymE2E at 2% of the cost — mustafasuleyman · 2026-11-18
- GitHub Copilot Fusion tops Terminal-Bench at 71.2%, costing 57-88% less than frontier models — mariorod1 · 2026-11-18
- Microsoft-Decision-1.1 hits GA with 2x speed at lower cost — marlene_zw · 2026-11-18
- OpenAI officially launches GPT-6.2 Astra and Sol models — kieranklaassen · 2026-11-18