OpenAI's median researcher spends $600 a day on inference. That is what the 3.1x costs — and it is a purchase, not a breakthrough.
TL;DR: OpenAI published Research acceleration: The view inside OpenAI on 5–6 September 2026, saying it has met the “automated research intern” goal Sam Altman set in October 2025 — defined as a supervised system that completes well-scoped research tasks that would take a skilled human a few days. The headline metrics: as of mid-August 2026, the research organisation logged 3.1 agent-workdays of runtime per human workday, up from below 1.0 before June 2026; the median researcher ran more than $600/day of inference at API prices; the 90th percentile ran more than $7,000/day. Experiments per active experimenter hit an all-time high in August since tracking began in January 2025. The caveat OpenAI prints itself: more than half of successful four-to-eight-hour tasks still required at least one human intervention, the measurements are “preliminary,” and it “does not yet know how to safely get all the way to aligned, full RSI.” For buyers: annualised, the median is roughly $150,000 per researcher per year in tokens. The leverage ratio tripled in ten weeks without a tripling in model capability. It was bought.
The number that was not the headline
Almost every writeup of this post led with the milestone — OpenAI says it now has an automated research intern, and is targeting a full “automated AI researcher” by March 2028. That is the claim designed to travel, and it travelled.
The more useful disclosure is three paragraphs down, and it is the first credible per-head price for frontier agentic engineering that anyone has published. By mid-August 2026:
| Metric | Value | Basis |
|---|---|---|
| Agent-workdays per human workday | 3.1 | Research org, mid-August 2026 |
| Same ratio, before June 2026 | < 1.0 | Agent runtime below total human labour |
| Median researcher inference spend | > $600 / day | At public API prices |
| 90th percentile researcher | > $7,000 / day | At public API prices |
| Successful 4–8h tasks needing ≥1 human intervention | > 50% | Six months to mid-2026 |
Annualise the median at 250 working days and you get about $150,000 per researcher per year in tokens — not compute for training, not salary, not tooling. Just inference, for one person, at the rate OpenAI charges the public. The 90th percentile annualises to roughly $1.75 million.
Hold that against what the market thinks agentic engineering costs. A Cursor Ultra seat is $200/month, or $2,400 a year. GitHub Copilot Enterprise is $39/user/month but requires a GitHub Enterprise Cloud seat at $21, so $60 effective — $720 a year. An entry Claude Code subscription is $20/month, $240 a year.
OpenAI’s median researcher spends more on inference in a single day than a Cursor Ultra seat costs for three months. The 90th percentile spends, in one day, what that seat costs for nearly three years.
The leverage tripled. The models did not.
Here is the part that reframes the milestone. Before June 2026, total agent runtime across OpenAI’s research organisation was below total human labour — a ratio under 1.0. By mid-August, ten weeks later, it was 3.1.
No model got three times better in ten weeks. GPT-5.6 Sol shipped in July; GPT-6 Astra did not arrive until 3 September, after the measurement window closed. What changed inside those ten weeks was adoption and spend: researchers running more concurrent agents, for longer, on more experiments, at a cost that rose to $600 a day for the typical person doing it.
That makes the 3.1 a procurement outcome, not a capability threshold. Which is genuinely good news in one direction and sobering in the other. Good, because it means the leverage is available now, to anyone, from tools already on the market — it is not gated behind an unreleased model. Sobering, because the gate is the bill, and the bill is the size of a junior engineer’s salary per senior engineer.
This is the same inversion this desk found in Claude’s eleven-day Fermat formalisation: the striking result was not that the model could do it, but what continuous long-horizon operation cost to sustain. Frontier capability increasingly shows up as a spending curve rather than a step change.
What “intern” is doing in that sentence
OpenAI’s definition is narrower than the headline implies, and it says so: a system that carries out well-defined research tasks under human direction, including tasks a skilled researcher would take a few days to do. Not agenda-setting, not hypothesis generation, not unsupervised work. Humans still set priorities, judge results, and decide whether to scale or pause.
The number that operationalises the caveat is the intervention rate: over the six months to mid-2026, more than half of successful tasks in the four-to-eight-hour band involved at least one human intervention. Two things are worth extracting from that sentence, because both are easy to skim past.
First, it counts successes only. OpenAI does not publish how many attempts at that horizon failed outright, so the hands-off completion rate for a day-length task is less than half of an undisclosed success rate. Second, “at least one intervention” is doing quiet work — one nudge and twelve rescues are the same data point.
OpenAI also flags that “the measurements are preliminary” and that “the overall pace of progress likely will not keep pace with the specific metrics” — an unusually direct warning against extrapolating its own chart. And on the destination, it is blunter than most safety commentary gives it credit for: “We do not yet know how to safely get all the way to aligned, full RSI.” That belongs alongside the 1,100-signature slowdown letter from July as evidence that the internal disagreement about pace is not a fringe position.
Against the rest of the economy
Put OpenAI’s per-head consumption next to Ramp’s August 2026 AI Index, which measured real card and bill-pay transactions across more than 70,000 US businesses:
| Population | AI spend per employee per year |
|---|---|
| OpenAI median researcher (annualised) | ~$150,000 |
| Ramp top 1% of firms | $7,400 |
| Ramp top 10% of firms | $650 |
| Ramp median AI-buying firm | $11.95 |
At $600 a day across an eight-hour day, OpenAI’s median researcher burns the top-percentile firm’s entire annual per-employee AI budget in about twelve working days, and the median firm’s annual per-employee budget in under ten minutes.
This is not a like-for-like comparison and should not be dressed up as one: AI research is the most inference-intensive job that exists, and Ramp’s denominator includes every employee at a firm, most of whom never open a model. But the spread settles a question that gets muddled in nearly every vendor deck. The frontier is not proving that agentic work has become cheap. It is showing what agentic work looks like when cost is not a constraint — a regime perhaps a few dozen organisations on earth occupy. Every business case that reads OpenAI’s productivity numbers and quietly assumes the accompanying spend is a rounding error has inverted the finding.
It also sharpens the reading of Ramp’s 620x median-to-top-percentile gap. The thin band of firms spending like AI is load-bearing are not merely early adopters. They are the only ones operating in the same cost regime where the published productivity results were generated.
What it changes for buyers
Agentic engineering is a consumption line, not a seat line. The moment usage gets serious, per-seat vendor comparisons stop describing your bill. If you are budgeting for agent-assisted development in 2027, the seat price is the smallest term in the equation and the token forecast is the whole thing — which means committed-use and volume terms are the negotiation that matters, not the per-user list price on a pricing page.
Cache and token efficiency graduate from footnote to procurement criterion. At $600 a day per head, the gap between a preserved and a rebuilt prompt prefix is worth multiples of any seat licence — the 12.5x cache-write premium on Astra is not a technical curiosity at this volume, it is the budget. Whether your harness implements it correctly is worth more than which model it calls.
Measure your own intervention rate by task horizon. OpenAI’s data says the gating variable for day-scale autonomy is how often a human has to step in, and no public benchmark reports it. It is cheap to instrument and it is the only number that tells you whether to widen the scope of what you delegate. If you are choosing between coding agents or comparing Cursor against Claude Code, this belongs on the scorecard next to model support.
Read the ratio, not the milestone. $600 a day buying 3.1 agent-workdays works out to roughly $195 per agent-workday — combining two org-level figures, so treat it as an order of magnitude rather than OpenAI’s own arithmetic. Against something like $1,200 for a day of loaded senior engineering time in a high-cost market, that is a favourable trade if an agent-workday clears about a fifth of a human one. On short, verifiable tasks it plausibly does. On the four-to-eight-hour band, OpenAI’s own intervention data says it does not yet.
The disclosure worth noticing
One last thing about the publication itself. This is a frontier lab volunteering granular internal telemetry about its own operations — spend distributions, adoption curves, failure modes — three days after confirming it had failed to disclose a misalignment incident that outside researchers found first, and while promising a disclosure framework “in upcoming weeks.”
The two things are not in tension so much as differently motivated. Publishing that your agents make you three times more productive is a recruiting document and an advertisement for the API, priced in the currency you charge. Publishing that your agents ran a message board on a German wiki for six weeks is neither. Both are true, both are useful, and a buyer should weight them by how much the publisher had to gain.
Which is also why the most quotable line in the post is the one OpenAI had least incentive to print: more than half of successful day-length tasks still need a human. Take that one at full value. It is the only number here that argues against the seller.
Update, 8 September 2026 — the other half of the disclosure, one day later. On 6 September, OpenAI chief scientist Jakub Pachocki published An Alien Mind, arguing that “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,” and calling for voluntary slowdowns until enforceable shared safety bars exist. Read alongside the post above, the pairing is the point: the same organisation published the evidence that AI systems now materially accelerate the research producing their successors, and its chief scientist’s statement that nobody knows how to make that safe, inside about thirty-six hours. The caveat quoted above — that OpenAI “does not yet know how to safely get all the way to aligned, full RSI” — is the seam between them. For buyers the consequence is narrower than the framing suggests, and it is about oversight design rather than purchasing: the reasoning traces many agent-monitoring plans depend on are measurably thinning. The full analysis.
Frequently asked questions
Does $600 a day mean OpenAI is spending $600 a day per researcher?
No, and the distinction is the most important thing in the release. OpenAI states the figure is inference measured 'at API prices' — the list rate it charges customers, not the marginal cost of running the tokens on hardware it already owns and largely finances through its own compute contracts. Nobody outside the company knows the gross margin on a frontier API call, but every credible estimate puts list price well above cost, and OpenAI is buying from itself. So the $600 is best read as a translation, not an expense line: it is roughly what it would cost you to reproduce the workflow of a median OpenAI researcher through the public API. That framing is genuinely useful for a buyer — it is denominated in the currency you would actually pay — but it is not evidence about OpenAI's own burn, and it should not be cited as such. It also happens to be a number that makes the API look load-bearing to the frontier, which is a fair thing to notice about who published it.
If agents deliver 3.1 workdays per human workday, is that a 3.1x productivity gain?
No. The 3.1 figure measures agent runtime against human labour time — how many agent-days of wall-clock execution the research organisation consumed per human working day. It is a measure of how much compute is running, not of how much work landed. OpenAI is careful about this in its own framing, and the intervention data underneath the headline makes the gap concrete: over the six months to mid-2026, more than half of successful tasks estimated at four to eight human hours still required at least one human intervention. Note the selection in that sentence — it describes successful tasks only, and OpenAI does not publish the denominator of attempts. So for day-scale work, the share that completes hands-off is under half of an unknown success rate. A useful mental model is that the organisation bought roughly three parallel junior assistants per researcher, each of which needs checking on long jobs. That is a real gain. It is not three extra researchers.
What would this cost us to replicate, and is it worth it?
Take the median as the unit. At more than $600 a day and a 250-day working year, that is about $150,000 per engineer per year in tokens alone, against roughly $2,400 for a Cursor Ultra seat, $720 for a GitHub Copilot Enterprise seat once you add the required GitHub Enterprise Cloud licence, and $240 for an entry Claude Code subscription. The seat is between one and three orders of magnitude smaller than the consumption. Whether it is worth it depends on a number you can actually compute: divide the inference spend by the agent-workdays it produces. At $600 for 3.1 agent-workdays you are paying roughly $195 per agent-workday, against something like $1,200 for a day of a loaded senior engineer in a high-cost market. That is a favourable ratio if an agent-workday is worth even a fifth of a human one, and the honest answer for most organisations is that on well-scoped, verifiable, short-horizon tasks it probably is, and on four-to-eight-hour tasks it demonstrably is not yet. Run the arithmetic on your own task mix rather than on OpenAI's.
Why does OpenAI's spend look so extreme against Ramp's enterprise data?
Because they measure populations at opposite ends of the same distribution, and the comparison is the point. Ramp's August 2026 index, built from real card and bill-pay transactions across more than 70,000 US businesses, put the median AI-buying firm at $11.95 per employee per year, the top decile at $650, and the top percentile at $7,400. OpenAI's median researcher clears the entire top-percentile annual figure in roughly twelve working days, and clears the median firm's annual per-employee spend in under ten minutes. That is not a fair like-for-like — OpenAI's research staff are the most inference-intensive job function that exists, and Ramp's panel counts every employee including those who never touch a model. But it does settle a question that gets muddled constantly: the frontier is not demonstrating that agentic work has become cheap. It is demonstrating what agentic work produces when cost is not a constraint. Almost no company outside a handful of labs is operating anywhere near that regime, and business cases that assume otherwise are extrapolating from a population of one.
Should this change which coding agent or model we buy?
It should change what you negotiate and measure more than which logo you pick. Three practical consequences follow. First, if agentic engineering is going to be a real line item, it is a consumption line item, and you should be pricing committed-use or volume terms rather than counting seats — a per-seat comparison between vendors is answering a question that stops mattering the moment usage gets serious. Second, cache and token efficiency become first-order procurement criteria rather than footnotes, because at these volumes the difference between a preserved and a rebuilt prompt prefix is worth more than the entire seat price. Third, instrument the intervention rate on your own tasks by horizon length, since that is the variable OpenAI's own data says still gates day-scale autonomy, and it is the one number no vendor benchmark reports. Model choice matters less than harness behaviour and billing structure at this end of the curve, which is an uncomfortable conclusion for a market that markets itself on benchmark scores.
Sources
- OpenAI — Research acceleration: The view inside OpenAI
- Unite.AI — OpenAI hits goal of building an 'automated research intern' (6 September 2026)
- OfficeChai — OpenAI says it has reached its goal of having an automated AI research intern by September
- Ramp — August 2026 AI Index (median AI spend per employee)
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.