OpenAI and Ironclad Research Shows GPT-6 Astra Cuts Contracting-Agent Time Nearly in Half Versus Sol

OpenAI published research with legal-tech company Ironclad on advancing computer-use agents for complex contracting workflows, reporting concrete performance gains for GPT-6 Astra over GPT-5.6 Sol.

Details

• Benchmark results: across 11 real contracting tasks, GPT-6 Astra scored an average of 55.0% versus GPT-5.6 Sol’s 41.6% — a 32% relative improvement — while cutting average completion time by 48%, finishing in 19.2 minutes versus Sol’s 37.0 minutes
• Task types: tasks included setting up nondisclosure agreements, creating procurement approval processes, and updating a reusable legal clause — workflows that typically take experienced human users 30–40 minutes
• Training method: OpenAI used synthetic training tasks modeled on representative workflows and reinforcement learning to improve the models through practice and feedback inside hosted Ironclad environments
• Evaluation rigor: each task carried between 8 and 50 distinct evaluation criteria depending on complexity
• Key finding: reliable enterprise computer-use automation requires agents to track multiple business rules simultaneously across multi-step workflows, a capability OpenAI describes as still under active development

What happened next

This is a research-and-benchmark disclosure rather than a product launch, giving a concrete, numbers-backed look at how much GPT-6 Astra’s computer-use capabilities have advanced over the prior Sol generation on real-world enterprise tasks — complementing GitHub Copilot’s own computer-use rollout for desktop apps earlier this month as evidence that computer-use agents are becoming a measurable, competitive capability across the industry.