Illustration: LittleDot concept art, AI-generated.
When people talk about GPT-7, the first question is often how powerful it will be. I have another one: if it really is stronger, will everyday work with AI become better value?
Value is something to calculate after a task is finished. One cheap request followed by three failed retries and a person cleaning up the result may not save money. A more expensive call that handles a complicated task correctly could be a better fit. Looking only at each request's price makes it easy to miss the rework that follows.
In Fireside Alpha's notes on Boris Power's conversation, I noticed discussion not only of GPT-7 and GPT-8, but also of larger models guiding smaller ones. The next generation of research may not show up only in the most expensive, headline-grabbing model. Improvements to everyday models are worth watching too.
This is my expectation, not a claim that GPT-7 has announced a price or capability. Its actual cost, speed, and available usage limits are not yet known.
Not every task needs the most expensive model
Building a website involves very different kinds of work. Rewriting a button, reading a short log, organizing links, and judging a database migration plan do not require the same effort. Using one expensive model for every step can inflate the bill; choosing the cheapest option for a difficult problem can create more work later.
I hope future tools make this choice easier. Use a capable, responsive model for straightforward tasks and bring in a stronger model when the problem is genuinely hard. When switching, carry over the materials already read and decisions already made so I do not have to repeat the background.
But that choice should not happen silently. With a paid service, I want to know which model was actually used, how much it cost, and whether failed retries were billed. The model can help choose; the account should still make the bill understandable. Otherwise, at the end of the month I may see only a total without knowing what to change.
If I were building the tool, I would run common tasks and compare completion rates, retries, and the total time and cost to finish each one. Put those together and we can tell whether a small model is truly economical or whether its low per-call price only looks good in isolation.
That is why GPT-7 news also makes me curious about smaller models. If a stronger model can help everyday models learn to work more reliably, we may benefit without calling the largest model every day.
When the next generation arrives, I want to compare two bills: the usage record and the time I spent. If both go down while the work gets better, that is an upgrade worth celebrating. The model that saves me money might turn out to be the less conspicuous small one in the product list.