Products

End-to-end research tasks and trajectories covering source discovery, evidence synthesis, citation validation, and long-form analysis across open-web and professional knowledge workflows.

Rich, stateful environments for coding, computer use, research, and professional workflows — paired with tasks, trajectories, and reward signals for long-horizon training.

Evaluation systems that turn nuanced outcomes into reliable learning signals, combining executable checks with carefully designed criteria for quality, correctness, and process.

High-signal demonstrations and multi-step traces that show models how experts code, use tools, recover from errors, and complete complex workflows from start to finish.

Pairwise arena evaluations where trained human judges compare model behavior at scale, revealing differences in correctness, usefulness, reasoning quality, and preference that static scores miss.

Domain experts create and judge work in science, finance, law, engineering, and other high-skill fields where surface-level fluency is not enough.

Data that teaches models to understand and act across screenshots, documents, interfaces, images, audio, and video — grounded in realistic tool-use tasks.

Production-ready datasets, benchmarks, and environments available without a custom build, spanning coding, research, agent workflows, and core model capabilities.