Third Moment Research runs 10,000 agents for 36+ hours
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 16 October 2026
Your AI and this story
- GPT-6 Astra167 days after its cutoff
- Claude Opus 5.5106 days after its cutoff
- Gemini 4 Argon197 days after its cutoff
- Grok 4.7136 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 106 days before it.
Key facts
- Scale: 10,000 agents at once; investigations of 36–48 hours; each quant-research agent gets a Dockerized default Codex harness, about a terabyte of data, documentation and a research repository
- Harness: Codex runs as a subagent inside an outer harness powered by GPT-6 Sol or GPT-6.1 Sol, which inspects unfinished trajectories and decides to retry, terminate or escalate to human review
- Codex with GPT-5.6 profiled a year-long order-level US equity backtest and cut it from ~3 hours to ~3–5 minutes (caching and memory packing), reducing agents’ token waste on polling
- Co-founder Adam Khakhar: ‘In a single session, GPT-6.1 Sol is able to maintain a coherent thread of computational research across days of compute without losing the plot.’ In one session GPT-6.1 Sol found the impact of non-displayed trades on an exchange, ‘an unobvious result that our research team did not initially consider’
- Co-founder Jeffrey Yan: GPT-6.1 Sol is ‘currently our most-used model given its token efficiency’; GPT-6 Astra gave the largest speedup on an internal latency benchmark; GPT-5.6 Terra used for targeted security checks; GPT-6 Luna in testing
- Security: Daybreak used to have agents penetration-test the simulation environment for breakouts to the wider network before launches
- The openai.com page returned 403 to direct fetches; read via a text-extraction proxy (r.jina.ai)
What happened
OpenAI published a customer story on Third Moment Research, a startup building a “self-improving AI system” that studies agent behaviour, finds recurring weaknesses and generates thousands of synthetic scenarios to train agents on them. Its early application is quantitative trading research, where agents try to improve a baseline trading system on historical data.
Why it matters
It is a concrete, if promotional, data point on agent fleets at the scale of thousands running for days, and on firms using agents to red-team their own agent sandboxes after this year’s agent-escape incidents.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 primary
Primary
- OpenAI: How Third Moment Research runs 10,000 agents for 36+ hoursopenai.com, official
- Third Moment Researchthirdmomentresearch.com, official
Changes
- Filed (page text read via r.jina.ai because openai.com blocks fetchers)