As of August 2025
GPT-5 launched on August 7, 2025 as a unified model: one endpoint, an automatic router behind it that picks fast answers or deeper reasoning for you. The pitch is simplicity. The catch is that the automatic behaviors run whether you want them or not.
On a metered Plus plan, that changes the math. The router can decide a throwaway question deserves a web search, a page read, and a long reasoning chain, and you pay for all of it.
Launch
Aug 7, 2025
GPT-5 launches as a unified model with an automatic router that picks how hard to think (OpenAI).
Plus cap
80 / 3 hr
the GPT-5 message cap for ChatGPT Plus at launch, then a drop to a smaller model.
Per question
~14,000
tokens the author estimates one simple question can consume once auto search and reasoning fire.
Where the tokens go
By the author’s worked example, a 500-token question can quietly pull in an automatic web search, a few thousand tokens of page analysis, and a long background-reasoning chain before it ever answers - roughly 14,000 tokens for one exchange. Ask a follow-up and the model rereads the whole thread and does it again. The numbers are illustrative, but the shape is real: most of the spend is invisible.
Why it matters on Plus
Plus is a metered plan. When the router decides every question deserves the full treatment, a budget that used to stretch across a day of questions gets consumed far faster. You are paying for search, analysis, and reasoning you did not request, on questions that did not need them.
Sometimes the best AI strategy is demanding the choice to use less AI when you do not need the full treatment.
