After introducing GPT-6 Astra and its ability to use computers for professional work, OpenAI’s next goal is making agents more capable and efficient with specialized software for complex business problems. The research focus is training models to understand a company’s business rules, execute multi-step workflows, and verify that completed work matches original requirements. To accelerate this, OpenAI is partnering directly with a small number of software companies; its first partner is Ironclad, an AI contracting company.

Together they developed tasks that require agents to configure agreements, approvals, and reusable legal terms. A representative case is a legal operations team setting up a software purchase process: Finance may need to approve purchases above a threshold, Security may need to review certain requests, and Legal may need to review nonstandard terms. An agent must keep those requirements in view while moving through the product, correctly routing requests above and below the threshold; individual correct steps are not enough. Ironclad employees and OpenAI employees who use Ironclad identified 11 tasks across legal, commercial, and procurement work, including setting up NDAs, creating procurement approval processes, and updating a reusable legal clause so it reflects a requester’s selected jurisdiction. The source estimates each task would take an experienced user about 30 to 40 minutes.

GPT-6 Astra is OpenAI’s first frontier model trained on the Ironclad tasks. The evaluation used 8 to 50 criteria per task depending on complexity, allowing the team to see where models succeeded and fell short. Ironclad provided hosted software environments for practice, and OpenAI researchers built synthetic training tasks around representative workflows and applied reinforcement learning so models improved through practice and feedback. Compared against GPT-5.6 Sol, using the highest-scoring reasoning settings for each model (Max for Astra, High for Sol), Astra scored an average 55.0% versus 41.6% on the 11 tasks, while estimated average time per attempt fell from 37.0 minutes to 19.2 minutes – a 32% higher score and 48% lower time. An internal model used during Astra’s development scored even higher, 63.7%, and OpenAI plans to carry those gains into future models. In one comparison clip, Astra met about 94% of task criteria in about 20 minutes, while Sol met about 85% in about 32 minutes.

For Ironclad, the collaboration highlights a problem its customers know well: a contracting process must handle exceptions while preserving the rules a business depends on. If an agent loses track of a rule mid-task, that limits what a software company can confidently ask it to do. The post stresses that human oversight still matters and that a full contracting platform remains essential. As models improve, Ironclad can use them in more of its own products, with AI taking on more of the effort in complex contracting while controls stay in place. OpenAI is inviting a small number of software companies to work directly with its research and engineering teams on tasks agents cannot yet reliably complete. Applicants should provide a concrete example of the task, evidence of where the agent fails, and how success would be judged, along with deep domain expertise, a secure testing environment, and data that can be safely used for research.