SID-2 launches search agent swarms claiming 8x speed and 20x cost savings over frontier models

AIFlow_ML · x · 2026-11-18

SID AI has launched SID-2 for private data search: it trains search swarms to find documents, emails, research and memories, claiming more relevant results than Opus and Astra while being 8x faster and 20x cheaper, with over 1T tokens processed in production.

Their core argument is "the Bitter Lesson for search": traditional RAG pipelines degrade as datasets grow or questions harden, and unlike math or coding, search gains little from bigger models — it gains from looking at more documents. So SID starts with a small agentic model that reasons and searches within 300ms, scaling up to 24 coordinating agents doing hundreds of searches. A single agent has 256K context; swarms act as effectively infinite context. Reasoning effort and metrics are configurable, and it works with existing search backends with no reindexing or data movement required.

Original post →

More from coding & agent

coding & agent channel →