Third Moment Says It Runs 10,000 OpenAI Agents for Up to 48 Hours

OpenAI published the 10,000-agent account on October 14

OpenAI’s October 14, 2026 customer story says Third Moment Research runs 10,000 agents concurrently on investigations lasting 36–48 hours. It names GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol as reasoning models, Codex as the action harness, and Daybreak as part of security testing. The page also displays one trillion OpenAI tokens consumed in a single day.

This is not a new model or public API launch. It is a provider-published account of one customer’s private system. OpenAI supplies quotes from Third Moment’s founders and describes the architecture, but publishes no system identifier, customer entitlement, capacity reservation, price, service level, aggregate success rate or independent evaluation. The verified news is that the companies disclosed the operating pattern and scale; the reported outcomes remain customer claims.

The outer harness decides whether unfinished work continues

Third Moment says quantitative-research agents run the default Codex harness inside Dockerized environments with a terabyte of data, documentation and a repository containing feature engineering, GPU-supported training, trading strategies and simulation modules. Codex can plan, use tools, create objectives, spawn subagents and compact context as work continues.

For agents that stop before finishing, Third Moment places Codex inside an outer harness powered by GPT-6 Sol or GPT-6.1 Sol. The outer layer examines the trajectory and decides whether to retry, terminate or escalate to a human. That is more specific than simply increasing a model’s context window, but the story does not publish the decision prompt, retry ceiling, escalation rate, false-completion rate, checkpoint format or recovery results after process, network or model failures.

A 97% backtest reduction is one optimization result, not an agent benchmark

OpenAI says Codex analyzed recorded simulation trajectories, timestamps and the codebase after agents repeatedly polled a slow backtest. It identified caching and memory-packing changes and deployed a fix that reduced a year-long order-level US-equity backtest from approximately three hours to approximately three to five minutes. Comparing three hours with the midpoint of that range yields about a 97.8% reduction, consistent with the page’s displayed 97% figure.

The article does not provide the repository, input size beyond the broad workload description, cache state, hardware, profiling output, patch, correctness checks, number of repetitions or runtime distribution. It also does not separate how much improvement came from caching versus memory packing. Treat this as a customer-reported case result—not evidence that Codex cuts arbitrary backtests, agent runtimes or token bills by 97%.

Long duration and token volume do not establish useful autonomy

Third Moment says GPT-6.1 Sol can maintain a coherent computational-research thread across days and is currently its most-used model because it uses fewer tokens while delivering performance similar to Astra. The team also says agents found an unanticipated effect involving non-displayed exchange trades and improved a baseline trading system. OpenAI reports these statements without publishing the strategy, holdout design, statistical tests, transaction costs, leakage controls, deployment status or financial outcomes.

Concurrency, wall time and token consumption measure activity and capacity. They do not measure whether an investigation was correct, novel, reproducible, economical or accepted by a human expert. A one-trillion-token day could indicate productive parallel search, expensive redundancy or both. The evidence needed to tell those stories apart—completed tasks, accepted findings, severe errors, duplicate branches, idle polling, retries and cost per accepted result—is absent.

Published token rates are not the system’s total price

OpenAI currently lists GPT-6.1 Sol Standard short-context rates at $2 per million input tokens, $0.10 cached input, $2.50 cache writes and $10 output. Requests above 272,000 input tokens move the full request to higher rates, and Fast, Ultrafast, Batch, Flex and regional processing carry different multipliers. GPT-6 Astra’s documented Standard rates are higher. Those public rates establish a menu, not Third Moment’s bill.

The story does not break the reported trillion tokens into model, input, cached input, cache writes, reasoning, visible output, retries or processing modes. It also omits Codex and tool charges, data and GPU infrastructure, orchestration, storage, security testing and people. Third Moment’s own site currently asks prospective customers to get in touch for a managed OpenAI-compatible endpoint and publishes no price. Buyers should request total cost per accepted research outcome rather than multiplying one token rate by the headline volume.

Daybreak is an access program, not proof of containment

OpenAI’s story says Third Moment uses Daybreak to test whether an agent could escape its research environment and reach the wider network before simulations launch. The current Daybreak API guide documents organization and project approval, model permissions and explicit cyber-program selection through access_programs.cyber. It also says selecting a program does not grant access and does not guarantee that every safety check or prompt will succeed.

Neither reviewed OpenAI page publishes Third Moment’s threat model, test corpus, environment boundaries, findings, remediation rate, independent review, residual risk or runtime enforcement. Pre-launch penetration testing can improve assurance, but it does not replace least privilege, network controls, credential isolation, immutable audit records, runtime detection, spend limits and a tested kill switch. The stronger claim—that 10,000 autonomous agents were proven contained—is not supported by the public evidence.

Third Moment’s public offer is narrower than the customer-story scale

Third Moment’s current site markets a fully managed, OpenAI-compatible inference endpoint. Its stated process detects recurring failures on approved real traffic, lets the customer confirm which issues to address, builds evaluations, improves the model or harness, and ships a revised endpoint. It says Third Moment operates serving, evaluation design, reinforcement-learning post-training, harness optimization and versioning.

The reviewed page does not promise access to the quantitative-research environment, 10,000 concurrent agents, 36–48-hour sessions or hundreds of thousands of collaborating agents. It provides a contact route rather than self-serve provisioning and does not publish pricing, support, security documentation, data-use terms, regions, models, capacity or service levels. OpenAI describes hundreds of thousands of subagents as Third Moment’s roadmap, not current availability.

What teams building long-running agents should do next

Trial one bounded investigation with a frozen question, data snapshot, tool set and acceptance rubric. Set ceilings for concurrent branches, wall time, total tokens, cache writes, external calls and spend. Require each branch to emit a checkpoint, evidence links, unresolved assumptions and a termination reason. Score accepted findings, severe errors, redundant work, human interventions and full cost—not simply duration or task completion.

Constrain the runtime with per-agent identity, deny-by-default network access, scoped credentials, sandbox and egress controls, checkpoint recovery, deterministic approval for consequential actions, and a global kill switch. Wait when the provider cannot show account-specific access, capacity and commercial terms. Reject a scale-up when complete trajectories, correctness gates, incident evidence, rollback and cost per accepted result cannot be produced.

Disclosures

AccessAllGPT received no OpenAI or Third Moment account, credits, endpoint, output, data, trace, benchmark, security report, briefing, review, payment or compensation for this article. OpenAI and Third Moment Research did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank the models or system. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI, Third Moment Research or organizations cited. Publication-wide relationships are listed on the disclosures page.