Updates

Terminal-Bench tasks: Stock overview at the top: 500+ tasks by domain, language and task type

Terminal-Bench tasks: Three sample tasks that differ in language, field and size, each with its task statement

Multi-turn coding agent trajectories: Headline figures: thinking share of output tokens and model calls per user prompt

Multi-turn coding agent trajectories: Five session excerpts from both models, each with its hard parts marked

Long-horizon computer-use workflows: Seven sample tasks, each with its full sequence of key screenshots

Blender 3D scene building: All ten samples shown from request to finished scene, with the agent's own screenshots in order

Browser 3D game development trajectories: Each sample game now shows its rounds, a screenshot sequence and two trajectory excerpts

New website and catalog, in English and Chinese.

Terminal-Bench tasks: Published with three sample tasks

Multi-turn coding agent trajectories: Published with a session excerpt

Long-horizon computer-use workflows: Published with four sample tasks

Blender 3D scene building: Published with ten samples

Browser 3D game development trajectories: Published with three sample games