Third Moment says OpenAI’s stack runs 10,000 agents for 48 hours.
Third Moment says it runs 10,000 AI agents for 36-48 hours. Here’s what powers it, where gains came from, and what’s unproven.
Illustration made on Lumeta with GPT Image 2.5Make one like it
On October 14, 2026, Third Moment Research said it is running up to 10,000 OpenAI-powered agents at once, with individual research sessions lasting 36 to 48 hours. The announcement matters because it points to a new scale for autonomous research systems, especially in quant work and code-heavy environments. For teams building agent workflows, the immediate takeaway is that Third Moment is using a layered stack: Codex for acting inside the environment, GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol for reasoning, and Daybreak to test security before runs begin.
What Third Moment disclosed
Third Moment described a system built to improve agents over time by analyzing where they fail, generating synthetic scenarios around those weaknesses, and feeding the results back into future runs. In its current setup, the company says agents can plan work, spawn subagents, use tools, and keep investigations going across long sessions instead of stopping after a single prompt chain. It also said the operation can consume 1 trillion OpenAI tokens in a single day, and that current evaluation work includes GPT-6 Luna alongside the other frontier models in its stack.
Illustration made on Lumeta with GPT Image 2.5Make one like it
Why the numbers matter
The most concrete example in the announcement was a backtesting bottleneck. Third Moment said Codex reviewed recorded simulation trajectories, timestamps, and the codebase, then found a caching optimization and a memory-packing technique that cut one backtest from approximately 3 hours to approximately 3 to 5 minutes, which the company described as a 97% reduction. That kind of gain matters because long-running agents can burn tokens and clutter their context while waiting on slow systems. Third Moment also said GPT-6.1 Sol is currently its most-used model for token efficiency, while GPT-6 Astra delivered the biggest speedup on one live latency benchmark.
Illustration made on Lumeta with GPT Image 2.5Make one like it
What comes next
Third Moment said it uses Daybreak before simulations launch to test whether agents could escape their research environment and reach the wider network. That security step becomes more important as the company pushes toward hundreds of thousands of collaborating subagents rather than isolated long runs. For other companies experimenting with autonomous research, the news is not just that bigger agent counts are possible. It is that scaling seems to depend on four pieces working together: strong reasoning models, a tool-using harness, performance tuning inside the workflow, and automated checks that the environment is still contained.
