Third Moment Research describes using OpenAI models, Codex, and Daybreak to run large-scale quantitative research agents and test their environments.
In a case study published October 14, 2026, OpenAI describes how Third Moment Research uses its models and tools to conduct extended computational investigations. The company says it can run 10,000 agents concurrently, with research sessions lasting 36 to 48 hours. Its system uses GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol, and Codex for research and computation, while Daybreak is used to test the security of agent environments.
One early application is quantitative finance. Agents work inside a Dockerized Codex environment containing a terabyte of data, documentation, and research code for feature engineering, GPU-supported model training, trading strategies, and simulations. They test strategies against historical data, judge whether apparent improvements are meaningful or noise, and use results to choose subsequent experiments. OpenAI’s account says this process has improved a baseline trading system developed by Third Moment’s engineers.
The case study highlights a backtesting bottleneck that Codex helped address. A test across a year of order-level US equity data took about three hours, during which agents repeatedly checked for completion. After analyzing the code and workload, Codex identified caching and memory-packing changes that Third Moment put into production. The backtest then took approximately three to five minutes, reducing the time agents spent waiting and checking status.
Third Moment also places an outer harness around Codex to handle unfinished tasks: it can retry, stop a job, or send it for human review. For security, the team uses Daybreak to check whether agents could escape their research environment and reach the wider network before simulations begin. OpenAI reports that the company’s roadmap aims to expand toward hundreds of thousands of collaborating subagents and longer chains of experiments; that is a stated ambition, not a current operating scale. The case study also presents model and performance figures as company-reported results rather than an independent evaluation.