OpenAI reported that GPT-6 Astra averaged a 55.0% rubric score across 11 Ironclad contracting tasks, compared with 41.6% for GPT-5.6 Sol, with estimated time per attempt falling from 37.0 to 19.2 minutes. The collaboration with Ironclad, OpenAI's first software partner for this research, turned workflows like configuring NDAs and procurement approvals into training and evaluation tasks. OpenAI used reinforcement learning and hosted Ironclad environments, and says an internal model reached 63.7%.
No score is assigned. Sources and their independence are shown in the citation chain below.