OpenAI researchers might be spending >$4M/day on tokens (at API prices)

An analysis of OpenAI's September 22, 2026 GPT-6 Sol and Luna announcement estimates that OpenAI researchers could be spending more than $4.24 million per day on tokens at API prices, based on the company's reported figures of over $600 daily for the median researcher and over $7,000 at the 90th percentile. The estimate assumes 1,125 researchers — 25% of the 4,500 employees reported by the Financial Times in March 2026 — with daily usage following a lognormal distribution, and excludes non-researcher usage and training compute. The analysis notes the $4.24 million figure is not a proven lower bound because OpenAI said usage "exceeded" the thresholds and the headcount and usage observations come from different dates.

OpenAI researchers might be spending $4m/day on tokens at API prices Based on this quote from the GPT6 announcement https://openai.com/index/introducing-gpt-6-sol-and-luna/ : “Valued at API prices, daily token usage has exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile”. Estimated slop ratioWhere 0% is full human with absolutely zero AI involvement and 100% is the cause of great dismay.: 80% September 22, 2026 An estimate for 1,125 researchers , assuming 25% of 4,500 employees are researchers. This values token usage at API prices; it does not estimate OpenAI’s internal costs. 1Assumptions OpenAI reports daily token usage above $600 for the median researcher and above $7,000 at the 90th percentile, valued at API prices. \ 1\ ref1 The Financial Times reported about 4,500 employees in March 2026, as relayed by Reuters. \ 2\ ref2 Assume 25% are researchers , and their daily usage follows a lognormal distribution . Treat $600 and $7,000 as exact quantiles for this calculation. Neither the 25% share nor the distribution is reported by OpenAI. 2Calculation Let X be a researcher’s daily usage value, measured in USD. For a lognormal distribution: \ 3\ ref3 Here, Φ⁻¹ is the inverse standard normal cumulative distribution; Φ⁻¹ 0.90 ≈ 1.281552. The mean—not the median—is the quantity to multiply by headcount: The estimate excludes non-researchers’ usage and training compute. It is an expected daily value under the assumptions, not a measured company total. 3Usage and concentration Limits. The source says “exceeded,” not “equal to.” Its two thresholds do not establish the mean or the upper tail, so $4.24M is not a proven lower bound. The March headcount and later usage observations are from different dates. All three charts are calculations, not employee data. References OpenAI. Research acceleration: The view inside OpenAI. https://openai.com/index/research-acceleration-view-inside-openai/ September 6, 2026, §1. Median observation: mid-August. Repeated in Introducing GPT-6 Sol and Luna https://openai.com/index/introducing-gpt-6-sol-and-luna/ , September 22, 2026. Reuters. OpenAI to nearly double workforce to 8,000 by end-2026, FT reports. https://www.reuters.com/business/openai-nearly-double-workforce-8000-by-end-2026-ft-reports-2026-03-21/ March 21, 2026. Reported headcount: 4,500; 8,000 was a hiring target. Reuters did not independently verify the FT report. NIST/SEMATECH. e-Handbook of Statistical Methods , §1.3.6.6.9: Lognormal Distribution. https://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm Distribution, quantiles, and mean. Parameters and results calculated here.