OpenAI Teams with Ironclad to Train AI on Complex Contracting

Overview

OpenAI has announced a strategic research partnership with Ironclad, a leading provider of AI-powered contracting software. The collaboration aims to advance the capabilities of GPT-6 Astra, OpenAI’s latest frontier model, by training it on complex, real-world business workflows. This initiative focuses on teaching AI agents to navigate specialized software environments and execute multi-step professional tasks with greater efficiency and accuracy.

By working directly with Ironclad, OpenAI intends to bridge the gap between general model capability and specific enterprise needs. The partnership allows researchers to define high-value research problems derived from actual customer challenges in legal and commercial operations. This approach ensures that improvements in model performance translate directly into tangible benefits for professionals relying on automated contracting tools.

Defining Complex Contracting Workflows

The core objective of this collaboration is to move beyond simple task execution and enable AI to manage the intricate lifecycle of business agreements. Contracting processes often involve multiple stakeholders, including finance, security, and legal teams, each with distinct approval requirements. An AI agent must not only complete individual steps but also ensure the entire workflow adheres to company policies and regulatory standards.

Ironclad and OpenAI researchers identified 11 representative tasks spanning legal, commercial, and procurement domains. These tasks include configuring nondisclosure agreements, establishing procurement approval chains, and updating reusable legal clauses based on specific jurisdictions. The complexity of these workflows requires the agent to maintain context and verify that all conditions are met before finalizing an agreement.

To facilitate this research, Ironclad provided hosted software environments where models could practice these tasks. Researchers developed synthetic training data around these workflows and employed reinforcement learning to help models improve through iterative feedback. This combination of expert-defined tasks, detailed evaluation criteria, and simulated practice environments ensures that the models are trained on scenarios that reflect genuine business priorities.

Performance Metrics and Model Comparisons

Early results from the partnership highlight significant performance gains for GPT-6 Astra compared to its predecessor, GPT-5.6 Sol. The evaluation focused on two primary metrics: the average score against detailed rubrics and the estimated time required to complete each task. Astra was tested using Max reasoning settings, while Sol was evaluated using High reasoning settings to identify their respective peak performances.

Metric	GPT-5.6 Sol (High Reasoning)	GPT-6 Astra (Max Reasoning)
Mean Rubric Score	41.6%	55.0%
Estimated Average Time per Attempt	37.0 minutes	19.2 minutes

Astra achieved an average score of 55.0%, marking a 32% improvement over Sol’s 41.6%. Furthermore, the estimated time per attempt dropped substantially from 37 minutes to just 19.2 minutes, representing a 48% reduction in latency. These metrics suggest that Astra is not only more accurate but also significantly more efficient at handling complex contracting assignments.

In internal development phases, a model used in the training of Astra achieved an even higher score of 63.7%, indicating potential for further refinements in future iterations. Video comparisons of the models handling identical tasks revealed that Astra met approximately 94% of task criteria in about 20 minutes, whereas Sol met roughly 85% of criteria in 32 minutes.

Implications for the Contracting Industry

This collaboration underscores the growing importance of context-aware AI in professional services. Ironclad emphasizes that a successful contracting process must balance flexibility with strict adherence to business rules. If an agent loses track of a critical compliance requirement midway through a task, its utility is severely limited. Therefore, human oversight remains essential as these technologies evolve.

“For AI to be genuinely useful in complex areas of contracting, it has to do more than complete individual actions. Agents need to understand the full contracting lifecycle, including how business workflows connect while preserving the controls teams rely on.”

The partnership offers Ironclad a unique opportunity to integrate these advancements directly into its product suite. As models become more capable of managing end-to-end workflows, legal and business teams may see AI taking on a larger share of the effort involved in complex contracting. This shift could reduce manual overhead while maintaining the rigorous controls that organizations depend on for risk management.

For OpenAI, the Ironclad partnership serves as a blueprint for engaging with specialized software companies. By co-developing tasks and evaluation methods with domain experts, the company aims to ensure that its frontier models are equipped to handle the nuanced demands of enterprise environments. This strategy positions OpenAI to deliver more robust and reliable AI agents for complex professional work in the near future.