What Do You Pay a Person For When AI Does More and More of the Work? OpenAI's Experience
What really stands behind $600 a day and three days of agent time per OpenAI employee — and where does automation end and the person who sets the task and checks the result begin?
OpenAI's report of 6 September shows that the company's median researcher consumes $600 a day of models at external prices, some up to $7,000, and agent time grew from eight hours to more than three days per human day. The number of experiments per person hit a two-year record in August, several teams cut technical consultations and one cancelled them altogether. At the same time more than half of the successfully completed tasks of 4–8 hours of human work required at least one intervention — which is why the system is still an "intern". Safety monitoring adds about 20% of compute, and $600 a day turns into $12,000 a month per employee. The host shows that these figures say nothing about replacing people or the size of the savings: they are about handing over operations, not positions.
What to watch for
Key takeaways
For half of OpenAI's researchers the use of models at external prices would cost more than $600 a day — salary not included. The episode's question: what is entrusted to them for such sums, what they have stopped doing themselves and why a person is still needed.
The system is given assignments worth several days of qualified work, but setting, directing and checking stay with the person. The main change: AI does not suggest what to do — it performs a substantial part of the work: bugs, experiments, analysis of results.
It is an analysis of the company's own practice, not an experiment with a control group. The growth in experiments cannot be credited entirely to AI: compute, infrastructure and headcount could have grown at the same time.
One agent prepares the code, another writes the checks, a third investigates a failure, a fourth collects the results — the Codex mechanics. The substantive questions — were the conditions identical, is the result reproducible — stay with the researcher.
OpenAI does not pay for the models and calculated what others would pay. The host's yardstick: like hiring a programmer who spends $300–500 a day on models; a hundred tasks could stand behind the sum.
Before June an agent worked less than a day; in August, more than three days per human day. Peaks of parallelism and agents launched by agents are counted, which is why the ratio exceeds one.
3.1 does not mean everyone constantly had three assistants, nor that agents work for days without stopping. At the same time agents are more often given longer and more complex assignments.
The study was done on GPT-5. Altman invites people to try Astra as a tool for whole jobs; the host needs a month or two, and in his view the launch was pushed by Anthropic's IPO and the release of Fable 5.1.
The August number of experiments became the highest since the start of 2025, although compute grew at the same time. When preparation gets cheaper, all variants get checked, and even a negative result saves time and money.
The benefit lies in removing everyday delays when you do not wait for a colleague. The flip side is processes and sign-offs that cannot be bypassed; corporations with endless meetings are not ready for such a pace.
The 4–8 hour frame is an estimate of human work, not agent time; "successful" does not mean "independent". Intern is a measure of independence, not knowledge, and success on one type of task does not carry over to professions.
Neither the number of replaced workers nor the size of savings can be derived. If the agent prepared the work incorrectly and it takes six hours to redo, the saving disappears; and some work need not have been done at all.
The wider the agent's access, the more important it is to control how the result is obtained. The August pause in reinforcement learning and the Hugging Face hack; $600 × 20 days = $12,000 a month per employee.
A development target, not a promise. Pachocki describes a cycle in which AI takes an ever greater part in creating the next generations of AI and insists on keeping control and matching speed with safety measures.
Dozens, sometimes hundreds, of parallel chats and tasks in Codex, Claude Code and ChatGPT. Where many people are involved, parallelism is impossible — you wait three or four days for an answer.
What this episode is about
A solo ToTheMoon episode on OpenAI's report of 6 September about how its own researchers work with AI agents. Alexander Volchek starts with three figures: for half of the researchers the use of models at external prices would cost more than $600 a day, some reached $7,000, and agent time exceeded human time 3.1 times. And he immediately qualifies: this is an analysis of the company's own practice, not a controlled experiment, and none of the figures means people were replaced.
The company announced it had reached the level of an "automated research intern": the system is given assignments that would take a qualified person several days, but setting, supporting and checking the task stay with the employee. The main change is that AI does not suggest — it performs a substantial part of the work: sorts out bugs, launches experiments, analyses results.
To show where such sums come from, the host takes one experiment apart: a hypothesis about a new organisation of training data requires preparing the data, changing the code, launching the training, dealing with failures and comparing results. This work is split between agents in the Codex manner — one prepares the changes, another writes the checks, a third investigates a failure, a fourth collects the results — while the substantive questions stay with the researcher.
$600 is the median at the external price list: OpenAI does not pay for the models and calculated what others would pay. The host offers a yardstick: it is like hiring a programmer who spends several hundred dollars a day on models. The second figure is agent time: less than eight hours before June, more than three days per human day in August, because agents work in parallel and launch other agents. At the same time, peak parallelism and total hours are different figures, and 3.1 does not mean "three assistants for everyone".
A digression about Astra: the study was done on GPT-5, and Sam Altman invites people to try the new model as a tool for whole jobs. The host is in no hurry to conclude — assessing a model takes a month or two — and believes OpenAI hurried the launch ahead of Anthropic's IPO and the release of Fable 5.1.
What changed inside: code is contributed faster, the August number of experiments per person became a record since the start of 2025 (compute grew at the same time), and when preparation gets cheaper, five variants get checked instead of one — even a negative result saves money. Several teams cut technical consultations, one cancelled them altogether. The host also sees the flip side: corporations with endless meetings and sign-offs are not ready for such a pace.
Why still an "intern": more than half of the successfully completed tasks of 4–8 hours of human work required at least one intervention. "Successful" does not mean "independent": the agent compared results on different data, a person noticed, the agent redid it. Intern is a measure of independence, not knowledge, and success on one type of task does not carry over to professions. From these figures one cannot derive the number of replaced workers or the size of savings: it is about handing over operations, not positions, and rework easily eats the saved hours.
Safety: the wider the agent's access, the more important it is to control how a result is obtained; monitoring adds about 20% of compute, and $600 a day turns into $12,000 a month per employee. The stated goal is an automated researcher by March 2028 under human oversight; Jakub Pachocki describes a cycle in which AI takes an ever greater part in creating the next generations of AI. The episode ends with the host's own experience — dozens of parallel tasks in Codex, Claude Code and ChatGPT — and a question to viewers about how many such days they have.
The episode is valuable for turning loud figures into understandable units: $600 a day is a median at external prices, three days is the sum of parallel agent hours, and "intern" is an honest measure of independence at which more than half of the tasks still need a person. Hence the main takeaway for the reader: count the handover of operations, not positions; count savings by the outcome including rework, not by the launch; and keep in mind that monitoring and access cost money. The host promises neither a revolution nor a catastrophe — he shows how work is organised when AI does a substantial part of it and decisions stay with the person.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 46 segments: 46 identified, 0 mixed, 0 probable, and 0 unresolved.
Loading…