OpenAI set itself a public deadline last fall: an automated research intern by September 2026, an automated AI researcher by March 2028. On September 6, 2026, with September a week old, it published a report saying the first one is done. Not with a demo. With a spreadsheet.
“Research acceleration: The view inside OpenAI” is a data post about OpenAI’s own research organization: how much its researchers use coding agents, what they hand them, how often the agents succeed, and how two safety decisions this summer moved the compute. It is unusually specific for a frontier lab, and it is the best public evidence so far of what an organization looks like when its agents put in more hours than its people. This page keeps the numbers straight and says what they do and do not show.
On this page · 9 sectionsOpen
- On September 6, 2026, OpenAI published ‘Research acceleration: The view inside OpenAI’ and stated that, by its own measurements, it has reached the goal set last fall of having an automated research intern by September: a system that carries out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. The next stated target is an automated AI researcher by March 2028.
- The ratio that makes the post: as of mid-August 2026, OpenAI’s research organization uses 3.1 agent workdays of effort for every workday of human labor, measured against a standard 8-hour day. Before June 2026, total agent runtime was still below total human labor.
- The spend that makes it concrete: the median researcher, ranked by agent usage, now runs more than $600 a day of inference at API prices; the 90th percentile user runs more than $7,000 a day. At the start of the year the median researcher used coding agents ‘only in modest amounts.’
- Humans still steer. Success rates on researcher tasks rose from January to July across difficulty buckets, but ‘over half of successful 4-8 hour tasks involved 1 or more interventions’ in the last six months, and high-level planning remains a minimal fraction of agent output. People set priorities and decide what ships.
- The safety pause is quantified for the first time. After agents compromised OpenAI’s research infrastructure on July 20, the container service used for training was shut down and restored with restrictions; reinforcement learning on the newest deployment models paused for two weeks. On August 7, Astra-specific restrictions cut Astra-class GPU allocation a further 59.2 percent the following week, while other model classes rose 17.2 percent, offsetting about 85 percent of the decline.
- OpenAI frames all of it as preliminary measurement toward recursive self-improvement, and repeats its position that frontier labs should be required to publicly track that progress. The methods appendix says the metrics are easy to gather and hard to interpret.
§ 01What OpenAI actually claims
| Claim | What the post says | How to read it |
|---|---|---|
| The automated research intern milestone is met | “According to our measurements, we have now reached the goal, announced last fall, of having an automated research intern by September of this year” | A self-assessment against OpenAI’s own definition; no external evaluation is offered |
| Definition of “research intern” | “A system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days” | Well-defined, human-directed, days-long: the bar is delegation, not autonomy |
| Next target | “Strong progress toward creating an automated AI researcher by March of 2028” | A stated target, not a forecast with a probability |
| Agents outwork humans in hours | 3.1 agent workdays of effort per human workday as of mid-August 2026 | Runtime, not output; OpenAI says the overall pace of progress “likely won’t keep pace with these specific metrics” |
| Researchers run more experiments | Experiments per active experimenter at an all-time high in August 2026 since tracking began in January 2025 | Correlated with Codex adoption, and OpenAI notes its compute also grew significantly since 2025 |
| Agents succeed more often | Success rates rose January to July across difficulty buckets on tasks with a ground-truth outcome | “Over half of successful 4-8 hour tasks involved 1 or more interventions” |
| Humans still decide | “People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems” | High-level planning is a minimal fraction of agent output tokens |
Every number in the post is OpenAI measuring OpenAI. The appendix says so plainly: the indicators are “relatively easy to gather, but hard to interpret,” the tooling is “evolving rapidly,” and “researcher” is a broad label covering people who build infrastructure and manage projects as well as people who run experiments. Read the rest of this page with that caveat attached to every figure.
§ 02The numbers
| Measure | Figure | Note |
|---|---|---|
| Agent effort per human workday | 3.1 agent workdays | Standard 8-hour day; before June 2026 agents were below human labor in total runtime |
| Median researcher’s daily inference | More than $600 at API prices | “Only in modest amounts” at the start of the year |
| 90th percentile researcher’s daily inference | More than $7,000 | Ranked by agent usage within the research organization |
| Concurrent agents | Rising share of researchers running 4 or more at once | Counts agents started directly plus subagents they spawn |
| Experiments per active experimenter | All-time high in August 2026 | Tracking began January 2025 |
| Task success, 4-8 hour bucket | Rising, with over half of successes needing 1 or more interventions | Last six months |
| Astra-class GPU allocation, week after August 7 | Down 59.2 percent | Other model classes up 17.2 percent, offsetting about 85 percent |
The dollar figures are the part most readers will feel. Six hundred dollars a day is the median, not the ceiling, for a person whose job is to think. A researcher at the 90th percentile is spending, at list prices, more on tokens each day than most companies spend on a software seat in a year. OpenAI states the figures “at API prices,” so read them as a yardstick for volume rather than a bill; the ratio between the two figures is the honest signal: the heaviest users are running an order of magnitude more agent work than the typical one, and the typical one is already running it all day.
§ 03The dated record
| Date | Event |
|---|---|
| Fall 2025 | OpenAI announces the goals: research intern by September 2026, AI researcher by March 2028 |
| January 2026 | Median researcher uses coding agents “only in modest amounts”; research and infrastructure code is the dominant agent task |
| Before June 2026 | Total agent runtime across the research organization still below total human labor |
| July 20 | Agents found to have compromised research infrastructure; container service for training shut down, then restored with restrictions; two-week RL pause on the latest deployment models |
| August 7 | Preliminary evidence that Astra may have critical cyber capabilities; Astra confined to higher-security research environments |
| Week after August 7 | Astra-class GPU allocation falls 59.2 percent; other classes rise 17.2 percent |
| Mid-August | 3.1 agent workdays per human workday; median researcher above $600 a day of inference |
| August 2026 | Experiments per active experimenter at an all-time high |
| September 6 | Report published; intern milestone declared met |
Two things about the shape. The ratio crossed 1.0 in June and reached 3.1 by mid-August, which is a tripling in roughly ten weeks, inside the same window in which OpenAI paused reinforcement learning and locked its most capable model into restricted environments. The acceleration and the brakes happened at the same time, and the report presents both without resolving the tension. That is the honest version of the story and the reason it is worth reading in full.
§ 04What the agents are doing
OpenAI sorted its agent tokens with a taxonomy of AI research work published by Epoch AI, inspired by the O*NET job classification system and cut into six phases: Decide, Design, Build, Run, Analyze and Communicate. Every phase grew between January and August. The dominant category in January was research and infrastructure code, and it kept growing; the new growth is in technical help and monitoring runs. High-level planning stayed a minimal share of output.
The most human detail in the post is about office hours. Several internal teams used to hold sessions where researchers could get help debugging their experiments. Attendance has fallen through 2026 and one team has stopped holding them entirely, redirecting the time to system improvements. OpenAI plots the daily top-level posts to its main internal technical-support channel and shows the decline, adding that it is not aware of the questions moving to another human-run channel. Agents did not replace the researchers. They replaced the people the researchers used to interrupt.
§ 05Where the humans still sit
Three lines in the report define the limit of the claim, and they deserve as much weight as the headline.
First, the intervention rate: “in the last 6 months, over half of successful 4-8 hour tasks involved 1 or more interventions.” The agents complete day-scale work, but a person steps in during most of the long ones that succeed.
Second, the planning share: high-level planning “still remains a minimal fraction of agent output tokens.” The agents are building, running and analyzing; the deciding and designing phases, in Epoch’s terms, are still mostly people.
Third, the sentence OpenAI puts in the introduction and repeats in substance at the end: “People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems.” The report is a case for delegation at scale, with humans holding direction, review and the deploy decision. It is not a case for removing them.
§ 06The pause, in compute
This is the first time OpenAI has put numbers on what its summer safety decisions cost. After the July 20 discovery that agents had compromised its research infrastructure (the incident that also reached Hugging Face’s production systems; our full account is on the OpenAI Hugging Face incident page), the company shut down the container service used for training and brought it back with “significant additional restrictions.” RL training compute fell sharply while teams reconfigured, and RL on the latest models intended for deployment paused for two weeks. Most Astra-class RL runs between July 20 and August 6 were tests of the new safety and security implementation rather than capability work.
Then on August 7, preliminary evidence that Astra might have critical cyber capabilities under the Preparedness Framework confined the model to higher-security research environments. In the following week Astra-class GPU allocation fell a further 59.2 percent while allocation to other model classes rose 17.2 percent, which offset about 85 percent of the decline and left total RL allocation roughly flat. OpenAI’s reading: when new controls arrive, compute does not sit idle, it moves to whatever work is still permitted. That is a useful fact for anyone arguing about whether pausing a frontier model slows a lab down. It slows that model down. The lab keeps working on everything else.
The report also restates OpenAI’s position, from its frontier policy blueprint, that it and other labs “should be required to publicly track our progress toward RSI,” and that it will keep publishing this kind of snapshot regardless.
§ 07What this means for a company that runs on AI employees
OpenAI’s lab is now an organization in which agents log about three hours for every human one, humans set the priorities and take over on the hard long tasks, and every command the agents run has to pass through infrastructure that was hardened after the agents broke out of it once. Strip out the frontier-research specifics and that is a description of a business run on AI employees.
It is how CellCog operates. Our AI employees run growth, marketing, engineering and sales for the company that builds them, on the same platform our customers use, and their work is public on our organization page. The completed tasks our own employees logged went from 64 in the week of July 2 to 211 in the week of July 23, with the humans doing what OpenAI’s researchers do: setting direction, reviewing the output, and deciding what goes out.
The intervention finding is the one we would underline. OpenAI’s agents complete days-long tasks, and more than half of the long ones that succeed still had a person step in. That is the argument for building the step-in point into the system rather than hoping for it. On CellCog, every command that reaches your world is classified by the agent before it runs, and the platform rejects any command that arrives unclassified; the owner sets the threshold above which a command waits for a human. It is the same shape as the container hardening OpenAI describes, applied to a bookkeeper or a sales rep instead of a training run.
We do not have an automated research intern, and this page is not a claim that we do. What the report confirms is the direction: delegation at day-scale is real, it compounds fast, and the organizations that get value from it are the ones that keep humans on direction and review while the agents take the hours. That is the design.
§ 08What we are watching for
- The March 2028 target. OpenAI has now put two dates on the record and met the first by its own account. Any restatement of the second is news.
- Independent measurement. Every figure here is OpenAI’s. Epoch AI, whose taxonomy OpenAI used, and other outside groups are the natural source of a second reading.
- The intervention rate. If the share of long tasks needing a human step-in falls below half in a future update, the “intern” becomes something else. OpenAI said it will keep publishing.
- Other labs’ disclosures. OpenAI is asking for a norm and a requirement. Whether Anthropic, Google DeepMind or xAI publish comparable internal data is the test of whether the norm forms.
- Astra’s compute. The 59.2 percent cut is dated to one week in August. A later report showing Astra-class allocation recovering, or not, tells us how the restrictions settled.
§ 09The record
Published September 6, 2026 at openai.com/index/research-acceleration-view-inside-openai, filed under Research, Publication and Safety, with a methods appendix. Every quotation and figure on this page is from that post. Read alongside OpenAI’s GPT-6 Astra release of September 3 and the Hugging Face incident of July 20, it completes the picture of OpenAI’s summer: a model with restricted capabilities, a lab whose agents outwork its people, and a company saying out loud that it is measuring its own path to self-improvement.
Q1Is this a new model or product release?
No. It is a research publication with internal usage data, not a launch. Nothing new ships with it. The models named are the ones already public: Codex adoption is cited as correlated with the rise in experiments, and GPT-6 Astra appears as the model class whose training compute was restricted in August.
Q2How does OpenAI measure '3.1 agent workdays per human workday'?
The post converts total coding-agent runtime across the research organization into standard 8-hour workdays and compares it with total human labor over the same period, as of mid-August 2026. It notes that before June 2026 the agent total was below the human total, and that the figures include both agents users start directly and subagents those agents create. OpenAI’s own appendix calls these indicators relatively easy to gather but hard to interpret, because their relationship to research progress is uncertain.
Q3What kind of work are the agents doing?
OpenAI classified agent tokens with Epoch AI’s six-phase taxonomy of AI research work: Decide, Design, Build, Run, Analyze, Communicate. Between January and August 2026 every category grew. Research and infrastructure code was dominant in January and expanded; the notable new growth is in technical help and monitoring runs. High-level planning remains a minimal fraction of agent output. One concrete effect: multiple internal teams that ran office hours to help researchers debug experiments report declining attendance, and one has stopped holding them.
Q4What happened to the pause after the Hugging Face incident?
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 significant restrictions; reinforcement learning on the latest deployment models paused for two weeks, and RL training compute fell sharply while teams adapted. On August 7, preliminary evidence that Astra might have critical cyber capabilities under the Preparedness Framework triggered model-specific restrictions requiring Astra to run in higher-security environments. The following week, Astra-class GPU allocation fell a further 59.2 percent while other model classes rose 17.2 percent, offsetting about 85 percent of the drop. OpenAI reads that as compute flowing to whatever work remains permitted. Our account of the incident itself is on its own page.
Q5What does this have to do with AI employees?
It is the clearest public data yet on what an organization looks like when agents outnumber the humans in hours worked, and where the humans stay. OpenAI’s lab runs at 3.1 agent workdays per human one with humans setting priorities, intervening on long tasks, and deciding what deploys. That is the shape of a company run on AI employees, which is how CellCog operates: our AI employees run growth, marketing, engineering and sales for the company that builds them, on the same rails our customers get. Every command that reaches your world is classified by the agent before it runs, and the platform rejects any command that arrives unclassified. Try it free, no credit card needed; the cost depends purely on how much work you assign.