OpenAI case study: Third Moment Research runs 10,000 agents for 36+ hours

OpenAI has published a customer story about Third Moment Research, an AI lab that runs agents at a scale most teams only simulate: 10,000 agents running at the same time and working for more than 36 hours at a stretch.

According to the post, the lab's stack includes GPT-6 Astra and GPT-6.1 Sol, paired with Codex and Daybreak, the last of which is used for cybersecurity.

The two behavioural claims matter more than the raw numbers: the agents can hold focus for days, and they improve as they are used.

Sustaining 10,000 agents for over a day and a half moves the hard part away from single-request latency and toward orchestration, state management, failure recovery and cost control.

Cybersecurity is called out as one of the workloads, with Daybreak appearing alongside the models, which suggests long-horizon agents are being pointed at tasks that require continuous vigilance rather than one-off answers.

For the wider market, the interesting shift is the unit of work: an agent's useful life used to be a conversation or a single task, and this case pushes it to a span measured in days.

What to watch next is whether this pattern survives outside a showcase, and whether the lab or OpenAI publishes details on cost, reliability and how often humans had to step in.