Ep 144: Language-server retrieval costs more tokens than grep for most coding-agent tasks and…

Models & Agents — Video Edition · Aug 17, 2026 · 11m

Models & Agents Language-server retrieval costs more tokens than grep for most coding-agent tasks and rarely improves success rates. What You Need to Know: A new measurement study on Claude Opus 4.8, Sonnet 4.6, and Haiku 4.5 finds that LSP-based semantic retrieval increases token use by 6-118% on symbol localization while delivering no recall gains over simple grep. The work introduces a tokens-to-success metric and shows that an adaptive router keyed on task class and model strength is the only configuration that sometimes saves tokens. Builders should test task-specific retrieval routing…