OpenAI Says Third Moment Research Runs 10,000 Agents At Once
OpenAI published a customer story on October 14, 2026 about Third Moment Research, a startup it says runs 10,000 agents concurrently on research sessions lasting 36 to 48 hours. Everything below comes from that page.
Transcript
OpenAI says a lab runs ten thousand agents at once. Its page reports they work for days.
The story describes agents tailored to a customer's own work. They improve with more use, it claims.
The article notes sessions can last thirty six hours. Recurring weaknesses become new practice scenarios, it adds.
The post explains that Codex found a slow backtest wasting tokens. Hours became minutes, it states.
OpenAI tallies a trillion tokens in a day. Software also tests whether agents can escape, it writes.
This is OpenAI's own customer story, and its figures are unverified. There is no product launch here.
Inboxsmith builds AI receptionists for small businesses. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Line by line backing for every narrated sentence, all drawn from the source page text and nothing else. Hook sentence one, OpenAI says a lab runs ten thousand agents at once: backed by the stat callout 10,000 Agents running concurrently and by the body line that Codex and the newer models support the system's research and computational work, enabling Third Moment to run 10,000 agents at once on investigations lasting 36 to 48 hours. Hook sentence two, its page reports they work for days: backed by the subhead phrase builds agents that focus for days and by the same 36 to 48 hour session callout. Scene one sentence one, agents tailored to a customer's own work: backed by Third Moment Research believes the best agents should be really good at that specific work, and by the preceding line that every company has its own workflows, tools, edge cases, and definitions of success. Scene one sentence two, they improve with more use: backed by should improve their capabilities the more they're used. Scene two sentence one, sessions can last thirty six hours: backed by the callout 36 to 48 Hours per autonomous research session. Narration speaks only the lower figure, written as words, and the full range stays on screen so no second number is voiced. Scene two sentence two, recurring weaknesses become new practice scenarios: backed by identify recurring weaknesses, and automatically create thousands of synthetic scenarios around those problems, and by the following line that the team uses those scenarios to improve agents, repeating the process as new issues emerge. Scene three sentence one, Codex found a slow backtest wasting tokens: backed by Codex uncovered a bottleneck, by the statement that backtesting a trading strategy across a year of order level US equity data took several hours, and by the line that while waiting, agents repeatedly polled for completion, wasting tokens and filling their context with status checks. Scene three sentence two, hours became minutes: backed by reducing the backtest from approximately 3 hours to approximately 3 to 5 minutes. It is stated qualitatively in narration so that no second figure is spoken in that scene. Scene four sentence one, a trillion tokens in a day: backed by the stat callout 1 Trillion OpenAI tokens in a single day. Scene four sentence two, software also tests whether agents can escape: backed by the Daybreak section, specifically the sentence that the tests check whether an agent could break out of its research environment and access the wider network, and by the statement that the team uses the Daybreak cybersecurity platform to automate that testing before simulations launch. Scene five makes no factual claim about the company; it is our standard disclosure, and it carries the required wording verbatim. Claims deliberately left out for lack of support: any trading return or profit figure, any dollar cost, any model pricing, any customer count, any headcount, any funding or investor detail, any claim of a launch, release, or availability, and any suggestion that the 97 percent figure was audited or reproduced. The page gives no baseline for that percentage, no date for the one trillion token day, no methodology for the concurrency count, and no description of what Daybreak actually found, and each of those gaps is named on a NOT PUBLISHED line on screen. Voice safety notes for the audit: exactly three numbers are spoken, ten thousand in the hook, thirty six in scene two, and a trillion in scene four, each in a different scene and never two in the same or adjacent sentences. No model version name is spoken. The source quote containing the idiom about losing the plot is not used in any field of this file.(Provenance note, not a real video title. The single source for this script is OpenAI's own customer story at https://openai.com/index/third-moment-research, titled How Third Moment Research runs 10,000 agents for 36+ hours, published and captured on October 14, 2026, the same day this Short was produced. The page sits inside OpenAI's customer stories template and carries the metadata fields Company size: Startup, Region: North America, Industry: Technology, and Products: API, Codex, closing with a Start building with OpenAI call to action. That makes it vendor marketing about a vendor customer. It is not a news report, not an engineering disclosure, and not a product launch, and nothing in it was reviewed by anyone outside the publisher and the customer. Because of that, every factual sentence of narration is attributed out loud to the publisher, a different attribution phrase is used in each sentence, and the final scene states plainly that OpenAI published this story about its own customer and its own products, that the figures come from that page, and that no outcome is independently verified. Corroboration: none exists. The research step searched YouTube for supporting material and retrieved zero of five candidate transcripts, because the installed yt-dlp build on this host cannot fetch captions. The dated timeline probes that came back were unrelated third party commentary videos about artificial intelligence in general, so no fact in this script rests on them and none of them is cited anywhere. No second outlet carried these figures at the time of writing, and no regulator filing, auditor report, or exchange record was available to check the runtime or trading claims. Every number shown on screen is a stat callout supplied by the page itself: 10,000 agents running concurrently, 36 to 48 hours per autonomous research session, a 97 percent reduction in backtest time with Codex, and 1 trillion OpenAI tokens in a single day. The two named human sources, Adam Khakhar and Jeffrey Yan, are co-founders of the customer company and are quoted inside the publisher's own page, so their statements are promotional testimony rather than independent confirmation. Region is listed as North America, so the regional screen is clean. No model version string is spoken in the audio; the version names appear only in onscreen text.)