Let the tokens flow. Sail powers efficient, reliable deep research

gist

Sail reports topping BrowseComp-Plus at 90.72% accuracy and $0.15 per query by treating deep research as a compute problem rather than a retriever-quality problem. A GLM-5.1 orchestrator never reads documents. Cheap gpt-oss-120b swarm readers ingest truncated hits in parallel, drop noise, and return compacted evidence, so the stack can raise k without blowing the main context. With simple Qwen3-Embed-8B/BM25 retrieval they match closed GPT-5 systems and beat the prior 68% open-source mark, because most of the 6.5B tokens land on efficient workers.

ideas

• The bottleneck is bulk parallel ingest, not a custom retriever. Single-agent plus SoTA retrieval fails; brute force would work with infinite context, so the real constraint is cheap compute that can raise k.
• Keep the orchestrator off the documents. GLM-5.1 only proposes searches, edits queries, and submits answers. Swarm readers filter and summarize so a retriever miss cannot flood the main context.
• Spend tokens on cheap workers. Across 830 queries the swarm consumed about 6.4B of 6.5B tokens. At 100 docs per search, readers are 99.4% of spend, which is why Sail quotes $0.15 versus $0.97 to $5.39 on other hosts.
• Simple retrieval plus a liberal k is enough. Qwen3-Embed-8B/BM25 plus parallel truncated readers reached 90.72% accuracy and 84.31% recall, matching GPT-5 plus proprietary retrievers (90.48%) and beating the prior 68% open-source mark.
• Read the cost table as infrastructure, not identical runs. Baseten prices used GLM-5, and the OpenAI and Z.AI rows used comparable models, so those multiples compare host prices rather than matched accuracy.

quotes