Study: Web Search Closes AI Model Performance Gap Significantly
PolarBearby · x · 2026-09-01
Braintrust evaluated 1,329 current events questions across 4 models and 14 conditions. Key findings:
- Gap Reduction: Search narrowed the accuracy gap between models from 47.9 points to 5.6 points.
- Recency Decay: Retrieval gain declined with event age, from 45 points for recent events to 24 for older ones.
- Diminishing Returns: Runs with 5+ searches scored 19–49%; a fifth query was often associated with lower performance.
- You.com Edge: For GPT-5.6 Terra, You.com search led built-in search by 3.66 points accuracy, with 3.40s faster latency and $0.0051 lower cost per question.
More from coding & agent
- Building a Private AI OS on 4x RTX 2080 Ti — askincihan · 2026-09-01
- Llama Mac app adds REST API request builder for llama.cpp — ggerganov · 2026-09-01
- Prototyping Tennis Game Level with Grok 4.6, Unity, and Blender — chongdashu · 2026-09-01
- Multi-Agent Memory Architecture: Targeted Retrieval vs. Unified Store — Fun-Following-1723 · 2026-09-01
- Vapi SDK Silent Failure: start() Resolves on Mic Block — skywave84 · 2026-09-01
- Solving Agent Detection and Memory Loss: Real Browser and Local Storage Solutions — Spliff_77 · 2026-09-01