Estimation question

Estimate the daily inference cost of running ChatGPT for its free-tier users.

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Structured estimation combining user scale, usage frequency, and compute cost per query.

State assumptions: assume roughly 200 million weekly active users, with free-tier users being the large majority, say 150 million.
Estimate daily active free users: assume 40 percent of weekly actives use it on a given day, roughly 60 million daily active free users.
Estimate queries per user per day: assume an average of 5 messages per active free user per day.
Estimate cost per query: assume a blended inference cost of roughly 0.01 to 0.02 dollars per message for a mixed-size model.
Multiply: 60 million users times 5 messages times an assumed 0.015 dollars equals roughly 4.5 million dollars per day.

Builds the estimate top-down from daily active users rather than starting from total registered users, which would overstate actual usage.
States the per-query cost assumption explicitly and acknowledges it varies a lot by message length and model size used for free tier.
Arrives at a clear daily figure, clearly framed as an order-of-magnitude estimate rather than a precise number.
Sanity checks the result against known infrastructure spend scale for large AI companies to confirm it is plausible, not off by orders of magnitude.

Using total registered or lifetime users instead of daily active users, which inflates the estimate significantly.
Skipping the per-query cost assumption and jumping to a final number without showing the arithmetic.

How would this estimate change if OpenAI moved free-tier users to a smaller, cheaper model?
What is the biggest lever OpenAI has to reduce this cost without hurting free-tier quality?