OpenAI’s own usage report puts its median researcher above US$600 a day in agent tokens. We think Canadian firms should read that figure as a ceiling on someone else’s economics and build their own AI agent token budget from the task up.
- OpenAI says its median researcher used more than US$600 a day of agent inference at API prices by mid-August 2026, and the 90th percentile used more than US$7,000.
- The same report says over half of successful four-to-eight-hour agent tasks needed at least one human intervention.
- Price an agent by the finished task, with review time included, before you set a per-seat allowance.
US$600 a day. That is what OpenAI says its median researcher was burning in coding-agent tokens by mid-August, priced at its own API rates, in a research acceleration report dated September 6, 2026. If your firm has started giving staff agent tools, someone in finance will soon put that number on a slide as the new normal. We think it is the wrong benchmark, and copying it would be an expensive mistake for most Canadian companies.
What does OpenAI’s US$600 a day actually measure?
It measures how many tokens a researcher’s agents consumed, multiplied by OpenAI’s public API prices. It does not measure what OpenAI paid, what the work produced, or whether the output was kept. The report says so in its own way: these figures are “relatively easy to measure, but can be hard to interpret.”
We agree with that line more than with the coverage that followed it.
Two other numbers in the report matter for a buyer. Agent runtime passed total human labour in June, and by mid-August the research organization was logging 3.1 agent-workdays per human workday. And on the tasks where OpenAI could check the outcome, “over half of successful 4-8 hour tasks involved 1 or more interventions.” That is OpenAI’s own grading (a vendor measurement, not an independent one).
Show Me the Invoice
Here is the median figure turned into a monthly bill, then set beside a more ordinary Canadian team.
| Figure | What it shows |
|---|---|
| US$600 | Median OpenAI researcher, agent tokens per day Fact, OpenAI report |
| US$7,000 | 90th percentile researcher, per day Fact, same report |
| 21 | Working days per month Illustration |
| 1.3784 | USD to CAD Fact, Bank of Canada, September 8, 2026 |
| US$200 a day | Canadian team: 10 staff at US$20 a day each Illustration |
- Median researcher: US$600 × 21 = US$12,600 a month, or about CA$17,368.
- 90th percentile: US$7,000 × 21 = US$147,000 a month, or about CA$202,625.
- Illustrative Canadian team: US$200 × 21 = US$4,200 a month, or about CA$5,789.
- Sensitivity: if that team’s usage runs ten times higher, the bill becomes about CA$57,893 a month.
The first two lines are a lab’s internal valuation. The last two are a real invoice, paid in Canadian dollars at a rate that moves.
Who captures the value when agents run all day?
Inside OpenAI, the value of extra experiments goes straight into the next model, which the company then sells. That loop is the whole point of the report, and it is why a frontier lab can justify a 90th-percentile researcher consuming six figures of tokens a month. A Calgary engineering firm or a Halifax insurer has no such loop.
Its agent spend has to come back as finished work a client pays for, or as hours someone stops billing internally.
The vendor captures the margin on every token either way. A buyer captures value only on the tokens that end in an accepted result. Tokens consumed and results accepted are two different counts, and early agent budgets tend to track only the first.
The bull case
The strongest counter-argument comes from OpenAI’s own data. Experiments per researcher hit an all-time high in August 2026, success rates rose across every task-length bucket from January to July, and one internal help channel saw traffic fall as agents took over troubleshooting. If those gains carry over to ordinary knowledge work, heavy agent spend could pay back fast, and a firm that rations tokens could fall behind one that does not. We take that seriously. Our view is that the evidence so far comes from one lab, measured by that lab, on its own research tasks.
What would change our view
An independent study, outside any AI vendor, showing cost per accepted task falling as agent spend per person rises would move us. So would a large Canadian employer publishing before-and-after numbers for a team on agents. Until then, we would set allowances by task and raise them when the numbers earn it.
Start from the controls we covered in Fable 5.1 pricing and cache reads, where a discount only helped bills of a certain shape, and from the questions in our shadow AI piece about who is already spending without a budget line. More cost analysis sits in AI in Business and our AI for business guide.
Your own numbers this quarter
Pull one month of agent or API usage for one team. Divide the spend by the number of tasks that were accepted without rework. Add review time at that team’s hourly cost. If the result is higher than what the same task cost before agents, hold the allowance where it is. If it is lower, raise it by a set amount and measure again next month. Repeat the check every quarter as part of your AI agent maintenance routine, since a model swap can change the cost per task overnight.
FAQ
How much do OpenAI researchers spend on AI agents?
OpenAI says its median researcher used more than US$600 a day of coding-agent inference at API prices by mid-August 2026, and the 90th percentile used more than US$7,000 a day.
What is an AI agent token budget?
It is a spending limit on the tokens an AI agent consumes, usually set per person, per team or per task, and reviewed against the work the agent actually finished.
Do AI agents still need human help on long tasks?
Yes, according to OpenAI’s own data. Over the six months to July 2026, more than half of successful four-to-eight-hour tasks involved at least one human intervention.
