Third Moment's 10,000 OpenAI Agents: Cost and Pricing

Third Moment Research runs 10,000 agents concurrently for investigations lasting 36–48 hours, according to OpenAI’s October 14 customer story. Its stack combines GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol with the Codex harness; Daybreak tests the environments before simulations launch.

The useful buying signal is not simply that agents can run for days. It is that infrastructure fixes, model routing and an outer supervisor can reduce wasted work. OpenAI announces no new model, endpoint, price cut or Third Moment service rate.

What the 10,000-agent story actually establishes

The quantitative-research environment gives agents a terabyte of data, documentation and modules for feature engineering, GPU training, trading strategies and simulation. Codex supplies planning, tools, subagents and context compaction.

An outer harness powered by Sol examines unfinished trajectories and decides whether to retry, terminate or escalate to a person. That is more specific than promising uninterrupted autonomy: completion still depends on orchestration and review.

Published claim	Useful evidence	Missing for a cost comparison
10,000 concurrent agents; 36–48-hour sessions	Customer-reported operating scale	Fleet utilization, failures and complete usage ledger
One trillion OpenAI tokens in a day	A headline scale figure	Model, billing-category, cache and service-tier split
Backtest falls from about 3 hours to 3–5 minutes	A concrete infrastructure optimization	Repeated trials, token savings and total dollar cost
Sol uses the fewest tokens with performance similar to Astra	Customer routing preference	Task distribution, quality scores and raw traces

The roadmap targets hundreds of thousands of collaborating subagents. That is an ambition, not the reported current deployment.

Why the 97% runtime claim is not a 97% discount

Codex with GPT-5.6 analyzed recorded simulation trajectories, timestamps and code. It found that agents repeatedly polled a slow backtest, spending tokens and filling context while waiting.

A caching optimization and memory-packing fix reduced that backtest from approximately three hours to three to five minutes. The story’s 97% headline describes backtest time, not end-to-end research time, model prices or an invoice reduction. No before-and-after token ledger is published.

Buyers should profile tools before increasing model spend. An event-driven completion signal and a faster simulator may free more research capacity than another reasoning upgrade.

Pricing impact: use the live rate card, not agent-hours

The table, chart and calculator on this page use our maintained canonical pricing dataset. The OpenAI pricing page explains Standard, cache, long-context and processing-tier rules; the token calculator lets you price your own workload.

Agent count multiplied by elapsed hours cannot determine an API bill. A waiting agent may generate few tokens; a busy subagent can generate many. Likewise, one trillion tokens does not identify a dollar bill without fresh-input, cached-input, cache-write, output, model and tier shares.

Above 272,000 input tokens, the full request uses the model’s long-context rates. Account for reasoning/output usage, retries, tool fees, Docker and GPU compute, data licensing, and human review. Batch or Flex discounts do not establish suitability for this interactive research loop.

Codex subscription allowances are not unlimited API inference. Daybreak’s controlled cybersecurity access is separate from ordinary model availability. Third Moment’s commercial terms are undisclosed; list prices do not reveal its negotiated bill.

Which model should research teams start with?

Third Moment says GPT-6.1 Sol is its most-used model because it completes tasks with fewer tokens while delivering performance similar to Astra. Astra reportedly produced the biggest speedup on a correctness-gated scoreboard optimization, but no numeric score or full benchmark package is supplied.

Use Sol as a complex-work control and Astra for harder investigations where the accepted result justifies the premium. Test Luna on narrow tasks; the story says Luna testing is ongoing, not that it matched Astra on deep research.

For a managed open-model control, compare available inference routes through Novita. Match tools, context and quality requirements; this is not a claim that Novita reproduces Third Moment’s stack.

Affiliate disclosure: we may earn a commission from the sponsored link at no extra cost to you.

Labs inclusion and explicit replay blockers

Our Labs leaderboard already includes exact GPT-6.1 Sol and GPT-6 Luna routes for a fixed 49-task text suite. It does not measure multi-day finance research. Astra remains unavailable on our maintained benchmark route; public API availability does not prove our account has access.

The coverage notes record the additional replay blocker: no released research dataset, pinned harness, scoreboard tasks, complete trajectories, baseline outputs, token bills or accepted-result grader. Daybreak also requires authorized cybersecurity access and a suitable environment. No customer claim becomes a Labs score.

What to do before scaling your agent fleet

Freeze representative tasks, data and success criteria. Run Sol and Astra with identical tools, reasoning settings and concurrency caps. Record tokens by billing category, retries, tool time, compute and reviewer minutes; reject invalid trading evaluations and data leakage.

Scale the route with the lowest total cost per accepted investigation—not the longest runtime or largest agent count. Third Moment offers useful orchestration evidence, but its public story cannot establish your production bill or investment returns.

Sources: OpenAI’s Third Moment Research story, API pricing, Astra documentation, GPT-6.1 Sol documentation, Luna documentation and Codex pricing. Sources checked October 14, 2026; API prices refresh through our daily pipeline.