GPT-6 Astra configures contract workflows more accurately: 55% of criteria met in Ironclad test

OpenAI trained Astra on tasks in the Ironclad contract platform and tested how well the model met the requirements for a finished workflow.

GPT-6 Astra met an average of 55% of the criteria in the Ironclad test published on Oct 6. This is the share of requirements met, not tasks completed.

OpenAI and Ironclad tested AI agents on configuring contract workflows within the software. For example, an agent must set up finance department approval for purchases and check that requests above and below a specified amount follow the correct routes.

Testing the finished workflow. Experts created 11 tasks, including nondisclosure agreements and purchase approvals. Each was evaluated against 8–50 criteria. Ironclad provided the working environments, while OpenAI trained models on synthetic tasks with feedback on the results.

In the research comparison, Astra met more requirements than GPT-5.6 Sol. The estimated time per attempt also fell.

| OpenAI measurement on Oct 6, 2026 | GPT-5.6 Sol | GPT-6 Astra | | --- | --- | --- | | Reasoning mode | High | Max | | Average share of criteria met | 41.6% | 55.0% | | Estimated time per attempt in simulation | 37.0 minutes | 19.2 minutes |

For your own automation. Computer use connects through the API: the model runs code to control a browser through Playwright or an application through PyAutoGUI. The application returns screenshots to the model and preserves the environment's state between calls.

OpenAI provides a sample application for this kind of experiment. The browser version requires Node.js 22.20.0 and an API key. In the project directory, run `pnpm install --frozen-lockfile`, `pnpm playwright:install` and `pnpm dev:js` in that order.

Original source: [OpenAI and Ironclad research from Oct 6, 2026](https://openai.com/index/advancing-computer-use-with-ironclad/). For implementation: [official computer use documentation](https://developers.openai.com/api/docs/guides/tools-computer-use) and [OpenAI's sample application](https://github.com/openai/openai-cua-sample-app).

OpenAI is accepting applications from companies with an example of an agent failure, success criteria, workflow experts and a testing environment.

GPT-6 Astra configures contract workflows more accurately: 55% of criteria met in Ironclad test

OpenAI trained Astra on tasks in the Ironclad contract platform and tested how well the model met the requirements for a finished workflow.

GPT-6 Astra met an average of 55% of the criteria in the Ironclad test published on Oct 6. This is the share of requirements met, not tasks completed.

In the research comparison, Astra met more requirements than GPT-5.6 Sol. The estimated time per attempt also fell.

OpenAI is accepting applications from companies with an example of an agent failure, success criteria, workflow experts and a testing environment.