Agent benchmarks

AssistantBench agent benchmark

OpenGPT keeps 15 published rows attached to the exact AssistantBench benchmark version and environment shown below.

Tasks: 33
Published rows: 15
Source data date: Aug 6, 2026
Identity reviewed: Historical

Evidence scope

HAL verified snapshot · 33 public tasks · 1 scaffold. public, automatically verifiable browsing tasks. Live-web research and browsing tasks reproduced by the Holistic Agent Leaderboard. Source-specific; no normalized step or token limit published.

How OpenGPT handles this evidence

OpenGPT preserves the publisher's metric, task scope, environment, version, configuration fields, and source date. It does not create a cross-benchmark overall Agent score.

Comparison boundaries

A higher result applies only to the displayed benchmark version and conditions.
Agent results reflect the full system, not the base model alone.
Missing cost, reliability, safety, or latency data remains unreported rather than being inferred.
HAL reproduced these rows and publishes traces, costs, and run counts. The leaderboard is a historical snapshot: HAL has paused adding new models, and reported costs do not include caching benefits.

HAL verified snapshot · 33 public tasks · 1 scaffold

Showing 15 of 15 published rows. Every result remains attached to the source benchmark and configuration.