OpenAI details Third Moment Research's 10,000-agent runs lasting 36-48 hours
OpenAI published a customer case study on how startup Third Moment Research runs a self-improving agent system on its models: 10,000 agents at once inside Docker environments, with individual investigations lasting 36 to 48 hours. The team says Codex working with GPT-5.6 analyzed thousands of recorded simulation trajectories and found that backtesting a year of order-level US equity data took hours; Codex then identified caching and memory-packing fixes and shipped them to production, cutting a backtest from roughly 3 hours to roughly 3 to 5 minutes. Third Moment wraps the inner Codex harness in an outer model driven by GPT-6 Sol or GPT-6.1 Sol that decides whether to retry, terminate or escalate to a human, and uses the Daybreak security platform for sandbox penetration testing; per 36Kr, the setup consumes about 1 trillion tokens a day, with GPT-6.1 Sol the most-used model for its token efficiency.
It is one of the largest publicly described long-horizon agent deployments, showing how supervisor models, parallel subagents and token efficiency shape the economics of running agents at scale.