OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
A research collaboration with the contract software firm produced 11 real-world legal tasks and a faster, more accurate model
OpenAI has a new study partner, and it brought homework from the legal department.
On October 6, 2026, OpenAI announced a strategic research collaboration with Ironclad, a company recognized as a leader in AI contracting. The goal is to change how AI agents are trained and evaluated on contracting workflows, the unglamorous paperwork that keeps businesses running.
The headline result: OpenAI’s GPT-6 Astra model scored an average of 55.0% across 11 contracting tasks. Its predecessor, GPT-5.6 Sol, averaged 41.6%.
The details: 11 tasks, half the time #
The two companies jointly defined 11 distinct tasks meant to mirror what contracting teams actually do all day. The list includes configuring agreements, managing approvals, and handling renewals of contractual elements. It also covers negotiation and jurisdictional approvals.
Ironclad contributed its domain expertise and had its own employees directly involved in defining what a successful result looks like.
Accuracy was only half the story. Speed improved too. With GPT-6 Astra, the average time per task fell from 37.0 minutes to 19.2 minutes. That works out to a 48% improvement in efficiency.
GPT-6 Astra itself is fresh off the assembly line. OpenAI introduced it in late September 2026 as its most capable model for professional tasks, so the Ironclad results serve as an early field report on that claim.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Background: a relationship that started with GPT-4 #
This partnership did not appear out of nowhere. Ironclad has worked with OpenAI products since 2023. That year, Ironclad launched its AI Assist product, which initially ran on GPT-4. The company has also integrated contract data into ChatGPT, bringing its customers’ agreements into OpenAI’s chatbot environment.
In 2026, OpenAI hired Ironclad co-founder Jason Boehmig, who now leads a new legal vertical at the AI company.
The research findings frame the effort as bringing customer needs and specialized expertise into the development of frontier models.
What this means for legal tech and beyond #
By letting a domain specialist define the tasks and the grading criteria, OpenAI is effectively outsourcing the question “what does good look like here?” to people who already know.
For Ironclad, the payoff is straightforward. The company gets early influence over how frontier models handle its core use case, which should feed back into products like AI Assist. With a dedicated legal vertical now led by Boehmig, OpenAI appears to be building out legal as a serious line of business rather than a side demo.
There are caveats worth watching. A 55.0% average score is a clear improvement over 41.6%, but it is not a level most general counsels would sign off on for unsupervised work. It is also worth noting that these benchmarks were co-designed by one company with one customer base. How well GPT-6 Astra generalizes to contracting teams with different processes, industries, or jurisdictions remains an open question.
Two years ago, Ironclad was plugging GPT-4 into a product feature. Now its co-founder runs a legal division at OpenAI, and its employees are helping define what success means for the next generation of models.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our