Global Tech News For people figuring out what to build with AI.
AI

OpenAI says GPT-6 Astra averaged 55% on 11 Ironclad contract tasks

Two business partners review a printed contract beside an open laptop; illustrative photo, not an Ironclad software screen.

OpenAI says GPT-6 Astra, its model for operating software through a computer interface, averaged a 55.0% rubric score on 11 contract-workflow tasks developed with Ironclad, compared with 41.6% for GPT-5.6 Sol. The result tests whether an agent can carry business rules through a multi-step process in specialized software, but it covers a small research task set.

A task might ask an agent to set up a software-purchase process: create an intake form and agreement template, route requests to Finance, Security or Legal under specified conditions, then check that requests on either side of an approval threshold follow the right path. OpenAI says each task was graded against 8 to 50 criteria and reports a mean rubric score. It does not publish the formula for combining scores across tasks, so 55.0% is a task-specific rubric result, not a general accuracy rate.

Ironclad employees and people who use the platform at OpenAI helped define 11 tasks across legal, commercial and procurement work, including nondisclosure agreements, approval flows and jurisdiction-specific contract clauses. Models practiced in hosted Ironclad software environments. OpenAI says researchers also used reinforcement learning, a feedback-based training method, on synthetic tasks derived from public SEC contract filings after filtering designed to remove personal information; it says no OpenAI customer data or internal contracts, and no nonpublic Ironclad customer data or contracts, were used. The announcement does not say whether the 11 evaluation tasks were kept separate from training, so it does not establish transfer to unseen workflows.

OpenAI compared Astra at Max reasoning with Sol at High, the respective settings at which each scored highest. It estimated 19.2 minutes for an Astra attempt and 37.0 minutes for Sol from assumed model processing speeds; it did not measure customer time savings. The company says results cover only 11 research tasks, not all Ironclad workflows, and the announcement describes research rather than a customer-product rollout.

Sources