I’m continuing to early-adopt the leading-edge of paid AI models, a Sisyphean hobby if ever there was one. Here’s my hot take after a month of using GPT-5 Pro vs. the lower-powered GPT-5 models (Thinking and Instant).
TL;DR
Need an answer in 5 seconds and don’t care too much if it’s wrong? Pick Instant.
Need deep strategy from first principles? Choose GPT-5 Pro.
Want one model with the best balance of what matters? Pick GPT-5 Thinking.
Lazy? Leave it on Auto and get what you get (likely Instant).
My current go-tos
OpenAI for almost everything except coding.
For coding, I’m using Claude Sonnet-4.5 on Cursor and really liking it.
Speed
5-Thinking usually replies in about a minute.
5-Pro often takes much longer, more like 10 minutes. The outputs are good, but it’s difficult to have much back-and-forth at that speed.
5-Instant is very fast, but outputs are noticeably worse.
Internet access
5-Thinking can browse; 5-Pro cannot. If your question depends on today’s reality, that’s decisive.
Quality
I sense that 5-Pro is smarter and gives better answers than 5-Thinking, but I humbly submit that I’m probably not smart enough to meaningfully tell the difference at this point. Both are excellent.
The “help Kevin think through any question he can come up with” benchmark (HKTTAQHCCUW Bench, for short) has been saturated.
5-Pro: Strategy formation, blank-sheet problem framing, multi-step plans (“create a business plan to do X”).
5-Thinking: Fast iteration, refinement, pulling in fresh facts, turning strategy into shippable details.
If I’m using both, I’ll start with 5-Pro to outline the architecture of the solution, then switch to 5-Thinking to refine the details.
Cost
5-Thinking ($20/month) is 10× cheaper than 5-Pro ($200/month).
Bottom line
5-Thinking and 5-Pro both give impressive, helpful answers to my questions. 5-Thinking is 10× faster, can search the web, and costs one-tenth the price of 5-Pro. For my use cases, 5-Thinking clearly delivers more value than 5-Pro.
A final suggestion
When you prompt ChatGPT (or Claude, or Gemini), be aware of which model is actually responding, even if it’s just understanding what the default model is. That one piece of information will make you much better calibrated on current capabilities.
