Updated
Updated · OpenAI · Oct 6
OpenAI Trains GPT-6 Astra on 11 Ironclad Tasks, Lifting Scores 32%
Updated
Updated · OpenAI · Oct 6

OpenAI Trains GPT-6 Astra on 11 Ironclad Tasks, Lifting Scores 32%

3 articles · Updated · OpenAI · Oct 6

Summary

  • GPT-6 Astra was trained on 11 Ironclad contracting workflows and scored 55.0% on OpenAI’s research evaluation, versus 41.6% for GPT-5.6 Sol.
  • Estimated time per attempt fell to 19.2 minutes from 37.0 minutes, as OpenAI used synthetic training tasks, reinforcement learning and Ironclad-hosted software environments to teach multi-step legal and procurement work.
  • The tasks covered configuring agreements, approvals and reusable legal terms across legal, commercial and procurement workflows that OpenAI estimates take experienced users 30 to 40 minutes each.
  • OpenAI said the partnership is the first in a broader push to work with a small number of software companies on specialized business software, while Ironclad said stronger models could expand AI’s role in contracting with human oversight still needed.

Insights

With GPT-6 Astra scoring only 55% on legal tasks, what hidden risks remain before AI can independently negotiate your next contract?
If OpenAI's new agents handle messy legal workflows, who takes the blame when a synthetic error creates a real-world compliance disaster?
As AI shifts from drafting text to executing enterprise workflows, will strict deterministic rules ultimately throttle its true autonomous potential?