Atlas

Benchmarks

← All benchmarks

ATLAS (Exa, 200-task subset)

Search · 2026-10-08

Exa's ATLAS asks an agent to return a table of every entity that meets a set of conditions, with 2 to 10 attributes each, so it must search wide to find the entities and deep to fill every cell. This row is the fixed seed-42 subset of 200 of ATLAS's 547 tasks, run by Exa with every model in its Scout agent with Exa search and page fetch for up to 25 tool-call rounds and an hour per task. A returned row counts only if the entity and every graded cell are correct; the score is row F1 averaged over tasks.

Top models (higher is better)

ModelScore
GPT-6.1 Sol69.5
Gemini 4 Argon53.8
Claude Sonnet 5.546.8
Muse Spark 1.342.7
Kimi K342.1
GLM-5.339.8
Grok 4.739.5
DeepSeek V4.1 Flash 101731.9
Loading Atlas data…