Advancing Computer Use with Ironclad: OpenAI Turns Contracting Workflows into Agent Training Tasks

On October 6, 2026, OpenAI published the results of a research collaboration with Ironclad, a company that builds AI contracting software. On the surface it reads as a case study about contract tooling. In substance it shows how OpenAI plans to train agents next. The question is no longer whether a model can operate a computer. The question is whether it can understand a company's business rules inside specialized software, execute a multi-step workflow, and verify that its work meets the original requirements. OpenAI says that its earlier GPT-6 Astra launch already showed how far its models have come at using computers for professional work, from preparing documents to testing websites. The next goal is to make agents more capable and efficient at using specialized software to solve complex business problems. To get there, OpenAI is partnering directly with a small number of software companies that understand these workflows best. Ironclad is the first.

Task design is the heart of the effort. Consider 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 short list into an intake form, document templates, approval rules and a record of the final agreement. An AI agent doing the same work has to keep those requirements in view from the first click to the last. It must configure Finance approval above the spending threshold, and it must check that requests above and below that threshold follow the right paths. Getting each step right is not enough. The finished process must work across the situations it was designed to handle. Ironclad employees, and people who use Ironclad inside OpenAI, helped researchers identify 11 tasks across legal, commercial and procurement work. They include setting up nondisclosure agreements, creating procurement approval processes, and updating a reusable legal clause so that it reflects the jurisdiction a requester selects. OpenAI estimates that an experienced user would need 30 to 40 minutes per task on average.

The evaluation and training design deserves a close look. Each task was scored against 8 to 50 criteria, depending on its complexity. That granularity shows which parts a model got right and where it fell short, rather than collapsing the result into pass or fail. Ironclad also supplied hosted software environments of its product, where the models could practice. OpenAI's researchers built synthetic training tasks around representative workflows and used reinforcement learning so that models could improve through practice and feedback. Three elements work together: tasks chosen with people who know the work, detailed criteria, and a place to practice. OpenAI argues that this combination keeps the improvement tied to the real-world tasks that matter most to customers. Our reading is that a setup built on live software, itemized rubrics and reinforcement learning sits closer to how enterprises actually accept work than a public benchmark alone does.

The reported results are specific. OpenAI compared GPT-6 Astra with GPT-5.6 Sol, each at the setting where it scored highest: Max reasoning for Astra and High reasoning for Sol. Across the 11 research tasks, Astra averaged 55.0% against 41.6% for Sol, a relative gain of about 32%. Estimated average time per attempt fell from 37.0 minutes for Sol to 19.2 minutes for Astra, a drop of about 48%. In a paired demonstration on one task, Astra met about 94% of the criteria in an estimated 20 minutes, while Sol met about 85% in an estimated 32 minutes. An internal model used during Astra's development reached 63.7% on the same tasks, and OpenAI says it aims to bring those further gains to future models. GPT-6 Astra is the first frontier model trained on Ironclad tasks.

These numbers call for restraint. The sample is 11 tasks. The evaluation is OpenAI's own research test, not an independent third-party benchmark. The times are estimated, simulated figures, not measurements from production use. The two models also ran at different reasoning settings. The fair conclusion is directional: targeted training on a specific piece of software can raise both completion quality and efficiency at once. For Ironclad, the exercise tests a challenge its customers know well, because a contracting process must handle exceptions while preserving the rules a business depends on. For the wider industry, the signal is clearer. Frontier labs are pairing with vertical software vendors to turn professional workflows into tasks that can be trained and scored. Whoever holds the best task definitions and the best practice environments holds the raw material for the next round of agent gains. Buyers who evaluate agents should ask for the same itemized criteria, and not settle for a single success rate.