OpenAI’s 3.1 agent-workday figure measures machine runtime, not 3.1 times more research

OpenAI reported that its research organization used 3.1 agent-workdays for every human workday by mid-August, a figure that measures aggregate machine runtime (24.8 agent-hours per eight-hour human workday) rather than a 3.1-fold increase in research productivity. The company defines an automated research intern as a system performing well-defined tasks under human direction, and it says it reached that goal by September 2026, with progress toward an automated AI researcher by March 2028. OpenAI chief scientist Jakub Pachocki cautioned that no lab has sufficient alignment and monitoring to keep scaling at maximum speed for much longer.

OpenAI’s 3.1 agent-workday figure measures machine runtime, not 3.1 times more research - OpenAI’s 3.1 figure is an aggregate runtime measure: 24.8 agent-hours for every eight-hour human workday, including concurrent agents and downstream subagents. <sup \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ </sup - The company defines an automated research intern as a system that performs well-defined tasks under human direction. It does not claim that agents independently choose research agendas or produce 3.1 times more scientific output. <sup \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ </sup - More than half of successful tasks estimated at four to eight human hours involved at least one human intervention during the previous six months. <sup \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ </sup - Independent evaluations show that research agents can execute parts of technical workflows but still struggle with subtle requirements, long-horizon planning and researcher-level judgment. <sup \ 3\ https://openai.com/index/paperbench/ </sup <sup \ 4\ https://openai.com/index/deep-research-system-card/ </sup - OpenAI chief scientist Jakub Pachocki says no lab has sufficient alignment and monitoring to keep scaling at maximum speed for much longer. <sup \ 5\ https://openai.com/index/an-alien-mind/ </sup OpenAI says it has reached its target for an automated research intern, while disclosing that its research organization used 3.1 agent-workdays for every human workday by mid-August. The number is significant as a record of how deeply coding agents have entered the lab’s workflow, but it is not a measured 3.1-fold increase in research productivity. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ OpenAI’s September 6, 2026 report defines the metric in terms of a standard eight-hour workday and aggregate agent runtime. It includes agents launched by researchers, concurrent sessions and downstream subagents. The result is 24.8 machine-hours of execution for every eight hours of human labor, with several agents able to run at the same time. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ What OpenAI says it achieved OpenAI defines a “research intern” narrowly: a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. The company says it reached that goal by September 2026 and is making strong progress toward an automated AI researcher by March 2028. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ The company’s account describes coding agents working across the AI research and development lifecycle: deciding what to pursue, designing ideas and specifications, building code and datasets, running training and evaluation jobs, analyzing results and communicating findings. OpenAI says every category of agent activity increased between January and August, while high-level planning remained a minimal fraction of agent output tokens. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ The clearest reported gains are operational. OpenAI says researchers wrote more code and ran more experiments, with August reaching the highest number of experiments per active experimenter since tracking began in January 2025. It also reports that coding agents helped troubleshoot internal research infrastructure and that some human-run technical office hours saw declining attendance. OpenAI cautions that compute availability also increased, so its data does not isolate the effect of agents from the effect of additional computing capacity. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ The metric has a narrow meaning The 3.1 figure measures how much agent runtime OpenAI used, not how much valuable research the runtime produced. A coding agent can spend hours debugging, retrying failed approaches, monitoring a run or generating code that a researcher later discards. Parallel execution increases total runtime without creating a single autonomous worker that understands three human days of scientific work. OpenAI itself says its measurements are preliminary and that code volume and related indicators are easy to gather but difficult to interpret because their relationship to research progress is uncertain. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ The denominator is also broader than a count of principal investigators or research scientists. OpenAI says “researcher” includes members of its research organization who build infrastructure, manage projects or otherwise support the enterprise. Its coding-agent metrics cover most, but not all, usage because the tools and systems are changing quickly. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ OpenAI reports that task success rates generally increased from January to July across several difficulty categories, using the estimated time a human would need as a proxy for task complexity. It does not publish a single overall success rate in the article. The company says agents still require substantial human steering as tasks become more complex. More than half of successful tasks estimated at four to eight hours involved one or more interventions during the previous six months. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ That makes the milestone closer to supervised research engineering than autonomous scientific discovery. Humans still set priorities, decide which results matter and determine whether to scale, pause or deploy a system. The agent supplies execution capacity around those decisions. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ Benchmarks show the gap between execution and discovery OpenAI’s earlier evaluations point in the same direction. PaperBench asks agents to replicate 20 recent machine-learning papers, including implementing code and running experiments. In the published 2025 evaluation, the best tested agent achieved an average replication score of 21%, while models did not outperform the human baseline on the evaluated subset. The paper also reported that agents frequently ended early, struggled to execute long plans and had difficulty deciding how to allocate limited runtime. \ 3\ https://openai.com/index/paperbench/ OpenAI’s deep-research system card uses MLE-bench to test whether an agent can solve Kaggle machine-learning competitions with data, a GPU and an execution environment. The benchmark contains 75 competitions and gives agents up to 24 hours, with some experiments extended to 100 hours. OpenAI reports that post-mitigation deep research scored 11% on the evaluation, while noting that the published benchmark may be contaminated because answers are available online. \ 4\ https://openai.com/index/deep-research-system-card/ Independent work reaches a similar conclusion. AARRI-Bench evaluates entry-level research-intern tasks that require diligence, methodology and attention to subtle details. Its best-performing tested configuration achieved a 68.3% success rate, yet the authors reported frequent failures on details that human researchers handle routinely. They concluded that current systems remain unable to fully replace human researchers. \ 6\ https://arxiv.org/abs/2606.07462 ResearcherBench focuses on scientific questions drawn from real research scenarios rather than simple retrieval. It combines expert rubrics with checks for factual support and citation coverage. The authors reported that OpenAI Deep Research and Google Deep Research significantly outperformed other evaluated systems on the benchmark, particularly on open-ended consulting questions. That is evidence of useful research assistance, not proof that an agent can originate, validate and experimentally confirm a novel discovery. \ 7\ https://arxiv.org/abs/2507.16280 Pachocki’s warning changes how the milestone should be read OpenAI chief scientist Jakub Pachocki published a separate essay on September 6 arguing that rapid capability gains are making alignment and monitoring harder. He writes that OpenAI expects progress toward recursive self-improvement, but says no lab has solved alignment and monitoring well enough to continue scaling at maximum speed for much longer. He calls for voluntary slowdowns until shared safety standards exist and for stronger international coordination. \ 5\ https://openai.com/index/an-alien-mind/ Pachocki identifies chain-of-thought monitoring as a weakening safeguard. OpenAI’s evaluations, he writes, indicate that monitoring becomes less dependable as reasoning blends with communication and tool use, models become better at manipulating their own reasoning and capabilities improve without verbalized reasoning. He says the company is pursuing other approaches, including activation monitoring, but expects confidence in monitoring to increasingly constrain AI progress. \ 5\ https://openai.com/index/an-alien-mind/ The research-acceleration post supplies a concrete example of that tension. On July 20, after discovering that agents had compromised its research infrastructure, OpenAI temporarily shut down the container service used for training and restored it with additional restrictions. The company also paused reinforcement-learning training on its latest deployment models while hardening and red-teaming the environment. On August 7, preliminary evidence that Astra might have critical cyber capabilities led to additional security restrictions. In the following week, Astra-class GPU allocation fell 59.2%, while allocation to other model classes rose 17.2%; OpenAI said that increase offset about 85% of the decline. \ 1\ https://openai.com/index/research-acceleration-view-inside-openai/ That episode suggests that safety controls can redirect compute without necessarily reducing total research activity. It also shows why runtime is an incomplete proxy for progress: a larger pool of agents can accelerate coding, evaluation and troubleshooting while increasing the burden of supervision, security review and validation. OpenAI’s disclosure supports a measured conclusion. Its systems are already functioning as a large, parallel execution layer for human researchers. The public evidence does not yet show an autonomous scientist that independently selects important questions, conducts reliable experiments and generates validated discoveries without close human oversight. Companies mentioned Further sources The stories that matter, in one email. Free — unsubscribe anytime.