OpenResearch Intelligence Index: first open benchmark for end-to-end autoresearch
BenBlaiszik · x · 2026-11-21
alphaxiv launched the OpenResearch Intelligence Index, billed as the world's first open-source benchmark for end-to-end autoresearch, asking which models are best for RSI, literature understanding, and running real experiments.
They spent $7.4k in tokens across 8 models, 6 providers and 3 evaluation tracks: scientific knowledge (20 hand-picked Terminal-Bench-Science tasks), post-training ability (6 novel post-training tasks), and literature relevancy/taste (a PaperFindingBench subset). Each model runs in its native harness inside OpenResearch. The index so far includes Opus 5.
More from coding & agent
- "Compiles" Isn't "Correct": Devs Push Back on a Common AI Coding Fallacy — gklambauer · 2026-11-21
- 55-85% of long Claude session usage was re-reading history — dev open-sources /flight-deck fix — PawelHuryn · 2026-11-21
- isometric-cities: coding agents turn real neighborhoods into pixel-art maps on Mac — zsakib_ · 2026-11-21
- codemap-search: single-binary Rust MCP indexes Grafana's 24k files in 13-16s, benchmarked vs ripgrep and 4 MCPs — Designer-Seaweed4661 · 2026-11-21
- Anthropic's MCP Code Mode Cuts Agent Tool Context From 150K Tokens to 2K (98.7% Savings) — blaizedsouza · 2026-11-21
- Dev builds OpenRouter-style POC routing tasks across Browser Use, Browserbase and TinyFish — Impossible-Comb4759 · 2026-11-21