OpenAI trains GPT‑6 Astra on Ironclad contracting tasks
OpenAI named contracting platform Ironclad as the first partner in a program that turns specialized business-software workflows into training tasks for computer-use agents, and says GPT‑6 Astra beats its predecessor on 11 Ironclad tasks.
OpenAI has started a research program with a small number of software companies to train and evaluate computer-use agents on specialized professional workflows. Its first partner is Ironclad, which makes AI contracting software.
Ironclad staff, along with OpenAI employees who use Ironclad, helped pick 11 tasks across legal, commercial and procurement work. Examples include setting up nondisclosure agreements, building procurement approval processes, and updating a reusable legal clause so it reflects the jurisdiction a requester selects. OpenAI estimates each task takes an experienced user 30 to 40 minutes. Each task is scored against 8 to 50 rubric criteria.
Ironclad also provided hosted copies of its product for models to practice in. OpenAI researchers built synthetic training tasks around representative workflows and trained on them with reinforcement learning. OpenAI calls GPT‑6 Astra its first frontier model trained on Ironclad tasks.
On the 11 tasks, Astra at Max reasoning had a mean rubric score of 55.0%, against 41.6% for GPT‑5.6 Sol at High reasoning, a 32% relative gain. Estimated average time per attempt fell from 37.0 minutes to 19.2 minutes. OpenAI notes that these times are estimates from simulation, not wall-clock measurements. An internal model used while developing Astra scored 63.7%, and OpenAI says it aims to bring those gains to future models.
The evaluation is small and internal, and OpenAI has not released the tasks or rubrics. The company said human oversight is still needed for complex contracting work. It is inviting other software companies to apply with concrete examples of tasks where agents fail, along with how success should be judged.