At the time of writing (I don't know if this is still the case), o1 was a paid model. This means that a technology that was free for 1+ years is now being sold for money by OpenAI, which to me personally seems like a scam. If they let everyone use it in the free tier, I wouldn't be writing this article.

Not only that, but it's much more restricted and limited than open source solutions.

Finally, a lot of people have tried to replicate OpenAI benchmarks, and o1 wasn't as good as described on the website. It is still better than other models, but if other models also used CoT technique to solve every problem, what would the benchmarks look like? I always thought of benchmarks as a test for LLM without any "modifications". But that's not the case with the o1 - it's not a pure LLM.