Are most people using a chainsaw to slice a pizza? You pick the effort, they pick the model. Who picks the bill?
Last week I ran a research task through ChatGPT: YouTube transcripts, web searches and S-1 filings. I turned the model up. Better answer. Tokens gone in short order.
Did I need it? I’m not sure. That is the point.
Three Key Things
1. Nobody knows which model they are using.
- The latest ChatGPT has dropped the model picker for a slider: instant, medium or hard.
- I choose the effort. They choose the model. I can’t see what sits behind each setting, or what it costs.
- Banks have the same problem at scale. In July, JPMorgan’s CFO told staff to match the model to the task: a report summary does not need the newest, most expensive model. He called first-half token spend trivial, but expects a meaningful rise in the second half.
2. Same pizza. Bigger chainsaw. Higher bill.
- The task does not change. Higher effort means the model thinks longer, and thinking is billed as output tokens.
- Past “good enough”, the answer barely improves. The cost keeps climbing.
- The catch: “good enough” moves with the task. A summary is a pizza cutter. S-1s plus transcripts plus search may need the chainsaw. Nobody shows you where the line sits.
3. Today’s price is the hook.
- Uber fares rose 92% from 2018 to 2021 once VC money stopped subsidising rides.
- Strava UK monthly went from £5.99 to £8.99.
- Google gives me Waze and Gemini for free. I’d pay for those.
- My price elasticity is low. That is what the vendors are counting on.
The counterfactual
- The better answer may be worth it. If the decision is worth thousands, a pound of tokens is cheap. The real cost is a wrong answer.
- AI keeps getting cheaper to run. A fixed level of capability falls in cost roughly 10x a year. Uber’s drivers never did.
Takeaway
Match the effort to the value of the answer, not the size of the question.
Something to think about: when did you last move the slider down?
Sources
- JPMorgan on model choice and token costs, Q2 earnings call, 14 July 2026: PYMNTS
- Inference cost trends: Wing VC; dsebastien notes
- Uber fares: Complex / Rakuten Intelligence
- Strava pricing: BikeRadar
Leave a Reply