Benchmark evidence

AssistantBench

AutoBot ranked #1 on the recorded official hidden-test leaderboard with 50.70% accuracy and a 100% answer rate across 181 tasks.

AssistantBench leaderboard with AutoBot highlighted at 50.70% accuracy

What was measured

AssistantBench evaluates difficult, time-consuming web research tasks. The reported result used AutoBot’s research harness with GPT-5.6 Sol Ultra and was scored by the official hidden evaluator.

Evidence

The repository records the method, submitted archive hash, official score, per-task results and leaderboard screenshot. The evaluation did not use hidden-label access, manual answer editing, ground-truth substitution or post-evaluation answer replacement.

Read the complete method

Leaderboard position can change as new systems are evaluated. This page describes the recorded result, not a permanent rank.