So we have language models and decision models. Don't forget search-models exist too. One way to make your Local AI agent faster might be simple, just read less of the web. Artificial Analysis tested Exa's new search setup with Dynamic Highlights. But instead of huge chunking of webpages into the LLM, Exa decides how much useful text to take from each page across one shared context budget. The results 📚 tokens/task → 299K → 157K 🔎 searches/task → 9.4 → 7.6 💰 cost/task → $0.083 → $0.066 And quality went UP. Artificial Analysis Search Index 🥇 Exa Auto + Dynamic Highlights → 80 🥇 Perplexity Search Medium → 80 On DeepSearchQA Exa → 85.6 F1 Perplexity → 80.8 When a local agent searches the web, every unnecessary token it gets back still has to go through your local model. So maybe we should also optimize what the model has to read → smaller 🎯 Exa itself is a hosted search API, not a local/open-weight search engine. Link in ALT