Third Moment says 10,000 agents run for 36 hours

No prediction venue prices Third Moment Research sustaining 10,000 agents for 36 to 48 hours, so there are no odds, no volume and no move to track after the OpenAI News profile described the system. That absence is the point, the claim is about harness and token efficiency, not a tradable outcome. Liquidity is zero. Not financial advice. Markets move fast.

OpenAI News reports Third Moment Research runs 10,000 agents at once, with investigations lasting 36 to 48 hours, and says a single day can consume 1 trillion OpenAI tokens. The lab frames that scale as a workflow shift for quantitative researchers and infrastructure teams who now have to budget for long-lived agent sessions, context compaction and automated penetration testing.

The budget question is immediate.

Adam Khakhar, co-founder of Third Moment Research, is the voice for capability in the profile. He said GPT-6.1 Sol can maintain a coherent thread across days without losing the plot, and he credited the GPT-6 Astra and GPT-6 Sol family with letting agents pursue root causes through hundreds of parallel subagents. He also said Codex acts as a subagent inside an outer harness powered by GPT-6 Sol or GPT-6.1 Sol that decides to retry, terminate or escalate to human review when a task stalls.

Jeffrey Yan, co-founder, is the voice for efficiency. He said GPT-6.1 Sol is the most used model because it completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra. That creates the tension in the piece. Khakhar points to Astra for the biggest speedup on a live scoreboard latency benchmark, Yan points to Sol for everyday use, and neither quantifies failure rates or human escalation volume.

For quant teams, the concrete change is in backtesting. The company said it used Codex with GPT-5.6 to analyze thousands of simulation trajectories and found agents polling during a three hour backtest across a year of order-level US equity data, wasting tokens and polluting context. After profiling, Codex applied a caching optimization and memory packing and cut the run to about 3 to 5 minutes, a 97% reduction according to the profile. Agents now get a Dockerized Codex harness, a terabyte of data and a repo for feature engineering, GPU training and simulation, with Daybreak used to test whether an agent could break out to the wider network.

The pricing context not in the profile helps explain Yan's preference. OpenAI said GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding and computer use at one-fifth of Astra's standard input and output prices, which frames why a lab running trillions of tokens would default to Sol for long sessions and reserve Astra or Terra for targeted checks. The profile also notes current testing includes GPT-6 Luna, without results.

No one in the source answered what these agents cost to run for 36 hours at scale, how often the outer harness escalates to a human, or whether the trading improvements cited have been independently validated. Until those numbers are disclosed, buyers have to price a system that improves with use on trust, not on a benchmark they can audit.