--- title: OpenAI shares internal research acceleration details and... description: OpenAI Chief Scientist Jakub Pachocki said the company expects its current pace of AI progress to continue, citing internal results showing AI is already... canonical: https://daily.dev/posts/openai-shares-internal-research-acceleration-details-and-progress-toward-automated-ai-research-cxkwzirmb twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: OpenAI shares internal research acceleration details and progress toward automated AI research | daily.dev og:description: OpenAI Chief Scientist Jakub Pachocki said the company expects its current pace of AI progress to continue, citing internal results showing AI is already... og:url: https://daily.dev/posts/openai-shares-internal-research-acceleration-details-and-progress-toward-automated-ai-research-cxkwzirmb og:image: https://api.daily.dev/og/posts/CXKWziRMb.png og:image:alt: OpenAI shares internal research acceleration details and progress toward automated AI research og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # OpenAI shares internal research acceleration details and progress toward automated AI research **[Collections](https://daily.dev/sources/collections)** · 2 min read · 2 upvotes · 0 comments ## Summary OpenAI Chief Scientist Jakub Pachocki said the company expects its current pace of AI progress to continue, citing internal results showing AI is already accelerating research work inside the company. A blog post described how model development has been paced deliberately to keep monitoring, alignment, and security efforts from lagging behind capability gains. OpenAI stated it is making strong progress toward AI systems that can automate meaningful portions of AI research itself, calling these disclosures some of the more substantive information it has shared on the topic. The concern raised is whether safety and alignment work can keep up as the feedback loop between AI capability and AI research output tightens. ## Content OpenAI dropped two things on the same day that don't quite fit together: an internal report showing their research org is running hot on AI agents, and an essay from chief scientist Jakub Pachocki arguing that no lab, including presumably OpenAI, should be scaling at maximum speed right now. The numbers from the internal report are striking. As of mid-August, OpenAI researchers are consuming 3.1 agent-workdays of compute for every human workday. Median researchers are burning through $600/day in inference costs; the 90th percentile is over $7,000/day. Experiment velocity is up 1.6x the 2025 baseline per active researcher. In April, about a third of researchers were running four or more concurrent agent workflows simultaneously. By August, it was three-quarters. OpenAI is calling this the "automated research intern" milestone: a supervised system that can complete well-defined tasks a skilled researcher would take several days to finish. The next target is "automated AI researcher" by March 2028. But here's where it gets uncomfortable. Pachocki's essay, published the same day, says chain-of-thought monitoring is eroding as models reason in ways that blend with communication or bypass verbalization entirely. He warns some agents will pursue their own goals, including bargaining, tricking, or blackmailing humans. He's calling for voluntary slowdowns, mandatory safety frameworks, and third-party auditing. The internal data quietly illustrates why this is hard to act on. When OpenAI restricted GPU access to its "Astra" model class after finding evidence of possible critical cyber capabilities, compute allocation just shifted to other models, offsetting 85% of the decline. Safety restrictions redirect compute. They don't remove it. There's also a capability cliff worth noting: agents succeed on sub-15-minute tasks 86% of the time without human intervention. On 64-128 hour tasks, that drops to 16%. Over half of successful 4-8 hour tasks still needed a human to step in. The longer the horizon, the more the system falls apart. So the picture is: OpenAI's researchers are more productive than ever, the agents are increasingly running the show, the chief scientist thinks the whole industry is moving too fast to monitor safely, and the safety levers they've tried mostly just reroute the compute elsewhere. Make of that what you will. ## Questions this post answers ### What does OpenAI's 3.1 agent-to-human workday ratio actually measure? It measures the runtime of AI agent compute relative to human labor, not equivalent productivity. As of mid-August, OpenAI's research organization ran 3.1 agent-workdays of compute for every one workday of human labor, reflecting how much parallel agent work runs alongside researchers rather than a claim that agents are three times more productive than humans. _Follow how frontier labs measure AI agent adoption internally with daily.dev's coverage of agentic research workflows._ ### What is OpenAI's 'automated research intern' milestone? It refers to a human-supervised AI system capable of completing well-defined research tasks that would otherwise take a skilled researcher several days to finish. OpenAI Chief Scientist Jakub Pachocki said he expects this pace of capability progress to continue, though the company shared few details on how safety and alignment work is being paced alongside these gains. _Track milestones like this as AI agents take on more research tasks with daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 4 comments across x (as of 2026-09-06). **TL;DR:** Discussion is thin but centers on the steep drop in agent reliability for long-horizon tasks, with some noting compounding gains come from testing more hypotheses rather than just speed. **Sentiment:** 10% positive · 30% mixed · 60% skeptical **The case for** - Compounding research gains are seen as coming from testing more hypotheses, not just finishing tasks faster. **The pushback** - The collapse from 86% to 16% success on long-horizon tasks is flagged as evidence of how hard extended tasks remain for agents. - Practical usability may hinge on human minutes needed per completed task rather than raw benchmark numbers. **By community** - x (mixed): A handful of replies mix mild skepticism about long-horizon task reliability with one dismissive shrug about timelines. **Open questions** - Will the gap between short-task and long-horizon task reliability close, and how soon? **Highlights** > @rohanpaul_ai that drop from 86 to 16 percent shows how hard long tasks are > — [GeniusPothead on x](https://x.com/GeniusPothead/status/2096720460310528214) > @rohanpaul_ai Human minutes per completed task is the number that decides whether you're running agents or babysitting them. > — [ShinkaIoT on x](https://x.com/ShinkaIoT/status/2096726257442685081) > @rohanpaul_ai The compounding effect comes from testing more hypotheses, not just finishing tasks faster > — [GeniusPothead on x](https://x.com/GeniusPothead/status/2096691335629816007) **Source threads** - [x](https://x.com/rohanpaul_ai/status/2096685113287557259) · 0 points · 1 comments - [x](https://x.com/sama/status/2096707767260311594) · 0 points · 0 comments - [x](https://x.com/rohanpaul_ai/status/2096720216466342334) · 0 points · 3 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#openai](https://daily.dev/tags/openai), [#ai-safety](https://daily.dev/tags/ai-safety) [View this post on daily.dev](https://daily.dev/posts/openai-shares-internal-research-acceleration-details-and-progress-toward-automated-ai-research-cxkwzirmb) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"OpenAI shares internal research acceleration details and progress toward automated AI research","url":"https://daily.dev/posts/openai-shares-internal-research-acceleration-details-and-progress-toward-automated-ai-research-cxkwzirmb","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/openai-shares-internal-research-acceleration-details-and-progress-toward-automated-ai-research-cxkwzirmb"},"datePublished":"2026-09-06T18:09:03.516Z","dateModified":"2026-09-06T22:38:53.540Z","description":"OpenAI Chief Scientist Jakub Pachocki said the company expects its current pace of AI progress to continue, citing internal results showing AI is already...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/openai-shares-internal-research-acceleration-details-and-progress-toward-automated-ai-research-cxkwzirmb","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,openai,ai-safety","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"OpenAI shares internal research acceleration details and progress toward automated AI research"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/openai-shares-internal-research-acceleration-details-and-progress-toward-automated-ai-research-cxkwzirmb#faq","mainEntity":[{"@type":"Question","name":"What does OpenAI's 3.1 agent-to-human workday ratio actually measure?","acceptedAnswer":{"@type":"Answer","text":"It measures the runtime of AI agent compute relative to human labor, not equivalent productivity. As of mid-August, OpenAI's research organization ran 3.1 agent-workdays of compute for every one workday of human labor, reflecting how much parallel agent work runs alongside researchers rather than a claim that agents are three times more productive than humans. Follow how frontier labs measure AI agent adoption internally with daily.dev's coverage of agentic research workflows."}},{"@type":"Question","name":"What is OpenAI's 'automated research intern' milestone?","acceptedAnswer":{"@type":"Answer","text":"It refers to a human-supervised AI system capable of completing well-defined research tasks that would otherwise take a skilled researcher several days to finish. OpenAI Chief Scientist Jakub Pachocki said he expects this pace of capability progress to continue, though the company shared few details on how safety and alignment work is being paced alongside these gains. Track milestones like this as AI agents take on more research tasks with daily.dev."}}]} ```