OpenAI and Ironclad contracting-agent research

OpenAI described a research collaboration with contracting-software company Ironclad on October 6. Ironclad staff and users at OpenAI identified 11 legal, commercial and procurement tasks, including setting up agreements and approval processes. OpenAI says its researchers built synthetic training tasks and used reinforcement learning in hosted Ironclad environments. In its internal evaluation, GPT-6 Astra scored 55.0% on average, compared with 41.6% for GPT-5.6 Sol under the reasoning settings selected for each model. These are company-reported results on a small, specialized task set, not an independent measure of general workplace reliability.