Skip to content
View original post on X: 空谷 Arvin Xu
Published
42/100

Interaction Bench 0.1 Preview ranks GPT-6 Astra first at 34.1%

AIInteraction Bench 0.1 Preview is a new benchmark that tests AI agents on 44 tasks drawn from 25 real open-source apps, with agents given only the user's goal and never the bug. GPT-6 Astra leads at 34.1%, followed by Claude Opus 5.5 at 27.3%, while the other models score between 4.5% and 15.9%.
空谷 Arvin XuVerified on X
@arvin17x
View on X

Introducing Interaction Bench 0.1 Preview — Benchmarked on what users feel

  • 44 tasks from 25 real open-source apps
  • GPT-6 Astra leads at 34.1%, Claude Opus 5.5 at 27.3%, others 4.5–15.9%
  • Agents get the user's goal, never the bug

GitHub: https://github.com/eval-sys/ibench
Website: https://ibench.ai

9🧵s next 👇

Source details

空谷 Arvin Xu · x.comPublished · added here

Original title: Introducing Interaction Bench 0.1 Preview — Benchmarked on what users feel

Topics

Open-source ecosystemModel releasesEvals & benchmarks