Advancing computer use with Ironclad
OpenAI describes a partnership with Ironclad to train and evaluate agents on contracting workflows, and GPT-6 Astra scored 55.0% against 41.6% for GPT-5.6 Sol on 11 research tasks.
OpenAI worked with Ironclad to turn 11 contracting tasks into research problems, each graded on 8 to 50 criteria. Ironclad supplied hosted software environments, and OpenAI used synthetic training tasks and reinforcement learning. GPT-6 Astra is the first frontier model trained on Ironclad tasks. Its estimated time per attempt fell from 37.0 to 19.2 minutes compared with GPT-5.6 Sol.
- Date
- Tuesday, 6 October 2026
- Lab
- OpenAI
- Kind
- feature
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Average score on 11 Ironclad research tasks, GPT-6 Astra (Max reasoning) | 55.0% GPT-5.6 Sol (High reasoning) scored 41.6%; the page also gives a 32% relative gain | company |
| Estimated average time per attempt | 19.2 minutes GPT-5.6 Sol took 37.0 minutes, which the page calls 48% lower | company |
| Internal development model score on the same tasks | 63.7% Internal model used in developing Astra, not released | company |
This is a research collaboration announcement and not a product launch. The access value is a closest fit, and the page names no new product. OpenAI is also inviting other software companies to partner.
Sources
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check