OpenAI tests computer-use agents on contracts—and the scores still leave room for oversight

OpenAI published an October 6 report on a research collaboration with Ironclad that turns contracting workflows into training and evaluation tasks for computer-using AI agents. The companies identified 11 tasks across legal, commercial and procurement work, including nondisclosure agreements, approval processes and reusable clauses. OpenAI says each task was evaluated against eight to 50 criteria in hosted software environments and that synthetic training tasks were derived from public SEC contracts after filters intended to remove personal information. The post says no OpenAI customer data or nonpublic Ironclad customer contracts were used for the evaluation described.

On OpenAI's internal research evaluation, GPT-6 Astra averaged a 55.0% rubric score versus 41.6% for GPT-5.6 Sol, while estimated time per attempt fell from 37.0 to 19.2 minutes. Those figures are vendor-reported results on 11 selected tasks, not an independent benchmark of every contract workflow or measured customer labor savings. OpenAI explicitly describes the times as simulated estimates. A 55% average also shows why improvement should not be confused with reliable autonomous completion: a system can become faster and score better while still missing requirements that matter in legal or procurement work.

The local relevance is not that Mountain House government or businesses use this tool; no such claim is made. The broader lesson applies to residents and small organizations adopting workplace agents: define the task, preserve approval rules, evaluate the final state and keep qualified human review where errors carry legal, financial or privacy consequences. The article does not provide legal advice or endorse either company. The contextual photograph shows a programmer working with code and does not depict the evaluated models, Ironclad software, contracts or customer data. Future coverage should look for independent replication, broader task sets and deployed safeguards.

How this was reported: Mountain House Live checked the linked OpenAI research publication directly on October 7, 2026, recorded the event or publication date separately from this edition's compilation time, and limited the article to details visible in that source. The accompanying image is licensed context, not evidence that the activity occurred or that a seat, service, decision or outcome is guaranteed. Readers should use the direct publisher link for late changes and current contact instructions. Where a source did not show a publication date, the source-date line says so instead of assigning one. No phone call, attendance, outcome or independent on-site observation is implied by the web verification. This distinction is especially important for future events: a calendar entry establishes a planned time and place, not turnout, quality, accessibility or completion. A future edition will repeat the topic inside seven days only if the responsible publisher posts a material, dated change; otherwise this version remains in the dated archive.

What comes next: before acting on this notice, reopen the official page, confirm the date, time, location, cost, eligibility and registration status, and use the publisher's contact route for individual questions. Do not send identity records, payment details, school files or health information to Mountain House Live. Public availability is not permission to reuse another person's photograph, recording or personal story. The article offers practical local context while keeping the authoritative decision, schedule and participation rules with the named agency. After the stated event date, the item should be treated as archive material unless the source issues a new date or a substantive follow-up. Corrections should preserve both the original source date and the date of the correction instead of making an old notice appear new.

Source-based account with Mountain House Live’s local analysis. Our interpretation is based on the linked records; it is not a statement from the source organization. Dates are preserved from this edition; confirm time-sensitive details with the original source.