Hey! I just found out that GPT-5.1 underperforms GPT-5 (same reasoning effort and sampling params). I wrote a x thread about it: https://x.com/Alex_Cuadron/status/1995633958957248716 Hope you find it interesting! 😄

Thank you, the results are very interesting! I'm afraid I didn't really understand what your hypothesis was, about why GPT-5.1 does worse than GPT-5 and even mini? 🤔 I'm curious, because I actually feel like it's worse on some things, but it's not easy to even have a good intuition as to why.

Thanks! IMO, GPT-5.1 is more aligned for the ChatGPT website and this caused it to perform worse in agentic benchmarks. Specially with a lower reasoning effort (lower than high), it seems like it was rewarded partially on the length of its responses (the shorter the better). So it has a tendency to hallucinate the answer (instead of checking against the environment) compared to GPT-5.