OpenAI trains Astra on Ironclad contracting workflows
OpenAI reports better results on 11 contracting-software tasks after training with Ironclad, while the evaluation remains internal and domain-specific.
Contracting tasks as a training ground
OpenAI says it has worked with Ironclad to make computer-using agents better at complex contracting workflows. Ironclad employees and users at OpenAI helped identify 11 tasks across legal, commercial and procurement work, such as configuring nondisclosure agreements, approval paths and reusable clauses. OpenAI’s researchers then created synthetic training tasks and used reinforcement learning in hosted Ironclad software environments. The company presents the work as a way to turn actual business rules and exceptions into training and evaluation problems; it is also inviting a small number of other software companies to collaborate on similar tasks.
What the reported scores establish
On OpenAI’s research evaluation, GPT-6 Astra averaged 55.0% across the tasks, against 41.6% for GPT-5.6 Sol. OpenAI says estimated time per attempt fell from 37.0 to 19.2 minutes. Each task was assessed against 8 to 50 criteria, and the comparison used the reasoning setting at which each model scored highest. These figures come from OpenAI’s own evaluation of a narrow task set, not an independent benchmark of autonomous legal work. The reported improvement still leaves many task requirements unmet. OpenAI and Ironclad describe human oversight and reliable handling of business rules as continuing requirements before agents can be trusted with consequential contracting processes.