Interesting article on the nature of “thinking” in deep reasoning models:
This has seemed intuitive to me. Both Gemini and GPT-5 employ much longer thinking cycles than Claude. With Claude, its 1-8 seconds mot of the time for me, with occasional stints up to 11 seconds or so. GPT-5 I regularly get thinking cycles a minute long or more, and Gemini is rarely less than 20-30 seconds.
Generally speaking, I have had much better results on most tasks with Claude’s models. GPT-5 has provided some improvement in certain areas, but in others it seems to be a right disaster (resolving testing issues, for example…ABSOLUTELY PITIFUL the last couple of days, while Claude is solving these problems left and right with ease it seems.
Further, I have also had the same kind of results with non-thinking Claude models, as with the thinking models, on many common tasks. A lot of coding tasks are faster and just as good with non-thinking. Some tasks do require some reasoning…solving these testing issues I’ve had for two days for example, seems to benefit from a thinking model.
Very interesting, though. A new kind of model is on the horizon: HRM or Hierarchical Reasoning Model. Curious to see if they really do bring a solid improvement to the field or not.