OpenAI Trains GPT-6 Astra on Ironclad Contracting Tasks, Cites 32% Higher Score

Training agents on specialized business software

OpenAI says its next goal is to make agents more capable and efficient at using specialized software to solve complex business problems. The company is exploring how to train models to understand a company's business rules, execute multi-step workflows, and verify that their work meets the original requirements.

To speed up this research, OpenAI is working directly with a small number of software companies that know these workflows best. Together, they identify challenging, high-value tasks and turn them into research problems for training and evaluating models. Ironclad, which OpenAI describes as a leader in AI contracting, is the first partner.

How contracting workflows became training tasks

OpenAI gives an example of a legal operations team setting up a process for buying software. Finance may need to approve purchases above a certain amount, Security may need to review certain requests, and Legal may need to review nonstandard terms. The person building the process must turn that list into an intake form, document templates, approval rules and a record of the final agreement. An agent doing the same job has to keep all of those requirements in view, for instance by configuring Finance approval above a spending threshold and checking that requests above and below it follow the right paths.

Ironclad employees and people who use Ironclad at OpenAI helped researchers identify 11 tasks across legal, commercial and procurement work. They included setting up nondisclosure agreements, creating procurement approval processes, and updating a reusable legal clause so it reflects the jurisdiction a requester selects. OpenAI estimates an experienced user would need about 30 to 40 minutes per task on average.

Each task was evaluated against 8 to 50 criteria, depending on complexity, which lets researchers see which parts a model got right and where it fell short. Ironclad also provided hosted software environments of its product where models could practice. OpenAI's researchers built synthetic training tasks around representative workflows and used reinforcement learning so the models could improve through practice and feedback.

What the results show

OpenAI compared Astra at Max reasoning with GPT-5.6 Sol at High reasoning, the settings where each model scored highest. Across the 11 research tasks, Astra averaged 55.0%, against 41.6% for Sol. Estimated average time per attempt fell from 37.0 minutes for Sol to 19.2 minutes for Astra.

OpenAI also reports that an internal model used in the development of Astra reached 63.7% on these tasks, and says it aims to bring those further gains to future models.

In one example task shown in two clips, Astra met about 94% of the criteria in an estimated 20 minutes, while GPT-5.6 Sol met about 85% in an estimated 32 minutes. These figures come from OpenAI's own research evaluation.

What it means for Ironclad and other software companies

OpenAI notes that a contracting process must handle exceptions while preserving the rules a business depends on. If an agent loses track of one rule midway through a task, that limits what a software company can confidently ask it to do. The company adds that human oversight still matters as agents improve, and that a full contracting platform remains essential.

For Ironclad, the collaboration is a way to bring customer problems into the development of the underlying model. OpenAI says that as models become more capable, Ironclad could use them in more of its own products, and that for legal and business teams this could mean AI taking on more of the effort of complex contracting while keeping the controls those teams rely on.

OpenAI is also inviting a small number of other software companies to work with its research and engineering teams on professional tasks that current agents still cannot reliably complete. It asks for a concrete example, evidence of where the agent fails, and a way to judge success. Partners should be able to bring people who know the work deeply, a secure testing environment, and data that can be safely used for research.