OpenAI has trained its GPT-6 Astra model to navigate specialized enterprise software through a research partnership with legal contracting platform Ironclad, marking an effort to push artificial intelligence agents beyond basic document generation into executing multi-step business procedures.
In benchmark tests across 11 tasks spanning legal, commercial, and procurement workflows, GPT-6 Astra achieved an average rubric score of 55.0 percent when evaluated using its Max reasoning setting. That performance represents an improvement over the older GPT-5.6 Sol model running on High reasoning, which scored 41.6 percent on the same evaluations. OpenAI reported that Astra completed tasks in an estimated average of 19.2 minutes per attempt, a 48 percent reduction from Sol’s 37.0 minutes. On one evaluated task, Astra met 94 percent of the requirements in an estimated 20 minutes, compared to Sol meeting 85 percent in 32 minutes. An unreleased internal development model reached 63.7 percent across the evaluation suite.
The research suite evaluated tasks that OpenAI estimates take an experienced human user 30 to 40 minutes to complete. These included configuring nondisclosure agreements, creating procurement approval chains, and updating reusable legal clauses to match specific jurisdictions. OpenAI evaluated each task against 8 to 50 criteria. Ironclad provided hosted software environments for the models to practice within, while OpenAI created synthetic training tasks derived from public corporate filings in the U.S. Securities and Exchange Commission’s EDGAR database with personal data removed. OpenAI trained the models using reinforcement learning with feedback inside these environments, excluding private customer contracts and internal agreements from the dataset.
The project targeted teaching agents to maintain business rules across chained actions rather than completing isolated steps. Sunita Verma, chief technology officer at Ironclad, stated that contracting workflows require agents to grasp the broader business lifecycle without bypassing institutional controls. Verma noted that systems must handle connected business workflows while maintaining the compliance checks that organizations implement to govern spending, legal liability, and security reviews.
Despite the measured score increases, the technology remains limited and dependent on human intervention. OpenAI noted that if an agent loses track of an organization’s business rules mid-workflow, it cannot be trusted to operate autonomously, making human oversight and contracting software safeguards necessary. The reported metrics reflect simulated time estimates based on assumed model speeds rather than measured customer time savings, and the results cover only the 11 designated research tasks rather than the complete range of Ironclad workflows. OpenAI is now seeking partnerships with other software developers to address professional tasks that frontier models still cannot complete reliably.
