Atlas

Benchmarks

← All benchmarks

PI-AutomationBench direct mode

Agents · 2026-11-03

Partial score of PI-AutomationBench in 'mode: direct', which registers each task's connected tools directly with the harness instead of leaving agents to discover them through search and execute tools. Prime Intellect's chart reports it with Prime Agent. A separate benchmark because the tool mode changes the task: DeepSeek V4.1-Flash 1017 gains 5 partial points and 18 strict points from it.

Benchmark source: Primeintellect

Top models (higher is better)

ModelScoreSource
DeepSeek V4.1 Flash 101796.8Prime Intellect, Learnings from Auditing Public Evaluations · Prime Intellect
Claude Opus 5.592.8Prime Intellect, Learnings from Auditing Public Evaluations · Prime Intellect
Loading Atlas data…