OpenAI's GPT-6 Astra scores 55% on Ironclad contract tasks

Now the fine print. The time savings weren't timed in anyone's office. They're simulated estimates based on assumed processing speeds. The tasks came from public contracts in the SEC's EDGAR database with personal details stripped out, and OpenAI says it used no private Ironclad customer contracts and none of its own customer data. That's good for privacy. It also means the agent hasn't practiced on the messy, one-off deals real legal teams actually argue over.

The test is small, too. It covers 11 research tasks rather than the whole product, picked with help from Ironclad employees and people at OpenAI who use Ironclad, and nobody outside OpenAI has checked the results. Each task was scored against a rubric of 8 to 50 criteria, and the 55% is an average of those rubric scores, so the model earned partial credit rather than a simple pass or fail on each task. On a set that small, one or two tasks going differently could shift the 13 point gap noticeably. An unreleased internal OpenAI model did better on the same set, so the version you can use isn't the company's best on this test. Neither side has said anything about pricing or exclusivity.

OpenAI says Ironclad is the first of a small group of software makers it plans to work with, and other companies can apply. The result worth waiting for will come from a live pilot: how often a person had to step in and fix the agent's work.