OpenAI and Ironclad turn contracting workflows into agent training grounds: GPT-6 Astra scores 55% on 11 tasks, about 32% above GPT-5.6 Sol

Why it matters

Key points

On October 6, 2026 (10:00 GMT), OpenAI described a research collaboration with Ironclad, the AI contracting software company, that turns real contracting workflows into tasks for training and evaluating agents on specialized software. Ironclad is the first partner in the program.

The setup has three parts. Ironclad staff and OpenAI employees who use Ironclad picked 11 legal, commercial and procurement tasks, such as setting up nondisclosure agreement workflows, building procurement approval rules, and making a reusable legal clause adapt to the jurisdiction a requester selects; OpenAI estimates each takes an experienced user 30 to 40 minutes. Each task is scored against 8 to 50 criteria. Ironclad also supplied hosted product environments where models could practice, and OpenAI built synthetic training tasks around these workflows and trained with reinforcement learning.

GPT-6 Astra is the first frontier model trained on Ironclad tasks. Across the 11 tasks, Astra (Max reasoning) averaged 55.0% versus 41.6% for GPT-5.6 Sol (High reasoning), and estimated time per attempt fell from 37.0 to 19.2 minutes. An internal model used in Astra's development reached 63.7%. OpenAI is now inviting a small number of software companies to join, asking for concrete tasks agents fail at, evidence of the failure, success criteria, domain experts and a secure test environment.

Verifiable

Primary page: https://openai.com/index/advancing-computer-use-with-ironclad (RSS pubDate Tue, 06 Oct 2026 10:00:00 GMT, category Company). Footnotes say results cover only these 11 research tasks, not all Ironclad workflows; times are simulated from assumed model speeds, not measured customer savings; training and evaluation used contracts from the SEC's EDGAR database with personal information filtered, and no OpenAI customer data or nonpublic Ironclad customer contracts. The task set, rubrics and full results are not public, and there is no independent replication.

Our take

The method matters more than the 32% gain. OpenAI is turning vertical software companies into suppliers of reinforcement learning practice grounds: the partner brings domain experts, task definitions, scoring criteria and a working product environment, and OpenAI brings training, so the model gets better inside that partner's software. Computer-use training has leaned on generic web and desktop tasks; collecting what doing the job right means, software by software and industry by industry, is closer to what enterprises actually pay for, and it gives whoever supplies good environments and rubrics a say in what the next models are good at.

The evidence is still thin, so it goes to Latest rather than this week's major changes. Eleven tasks is a small sample; Astra was trained on Ironclad tasks and then scored on similar ones, and outsiders cannot tell how close training and test are; a 55% average score is far from something you would let run unattended, and the times are simulated. Read alongside Astra for Law, it shows OpenAI treating legal and contracting work as an early proving ground for professional agents. Two things to watch: whether more software companies join and in which industries, and whether OpenAI turns industry tasks like these into public evaluations so outsiders can check how much progress is real.