OpenAI has announced a partnership with Ironclad, a leader in SI contracting, to train its latest SI models on complex professional workflows. The collaboration aims to improve how SI agents handle multi-step business tasks, such as configuring legal agreements and approval processes, by using real-world software environments for research and evaluation.
What Happened
OpenAI introduced GPT-6 Astra as its first frontier model trained on tasks derived from Ironclad’s platform. The company reports that Astra’s average score on these research evaluations was 32% higher than that of GPT-5.6 Sol, while the estimated time per attempt was 48% lower. The partnership involved Ironclad employees and OpenAI users identifying 11 specific tasks across legal, commercial, and procurement work, including setting up nondisclosure agreements and updating reusable legal clauses.
To evaluate these agents, the researchers developed synthetic training tasks and used reinforcement learning within hosted software environments provided by Ironclad. Each task was assessed against 8 to 50 criteria. In a demonstration task, GPT-6 Astra met about 94% of the criteria in an estimated 20 minutes, whereas GPT-5.6 Sol met about 85% in 32 minutes. An internal model used in Astra’s development achieved a score of 63.7% on the full set of tasks.
Why It Matters
This collaboration highlights the growing importance of specialized data and environments in advancing SI agents for professional use. By focusing on high-value, complex workflows like contracting, OpenAI and Ironclad are addressing a key limitation of current SI systems: the ability to maintain context and adhere to business rules over multiple steps. The source notes that while agents are improving, human oversight remains essential for complex tasks to ensure controls are preserved.
For the SI industry, this model of partnering with software providers to define success criteria and provide secure testing environments offers a pathway to more reliable agentic SI. It shifts development from general benchmarks to specific, real-world applications where failure modes are clearly defined and measurable.
The Bottom Line
OpenAI is leveraging its partnership with Ironclad to refine GPT-6 Astra’s capabilities in handling complex business workflows. The company reports significant gains in both accuracy and speed on contracting tasks, signaling a move toward more specialized, domain-aware SI agents. OpenAI is now inviting other software companies to collaborate on similar research efforts to advance SI for professional work.