It’s interesting that it one-shot it. I wonder if anyone had tried asking 5.4 Pro it before, and if so how often it produces this proof.
I recall that after 5.4 Pro solved the first Epoch open problem, Epoch found that Opus 4.6 and Gemini 3.1 could both also solve it.
GPT-5.4 Pro replicated this 8 times out of 10 when Arb Research tried it.
It’s interesting that it one-shot it. I wonder if anyone had tried asking 5.4 Pro it before, and if so how often it produces this proof.
I recall that after 5.4 Pro solved the first Epoch open problem, Epoch found that Opus 4.6 and Gemini 3.1 could both also solve it.
GPT-5.4 Pro replicated this 8 times out of 10 when Arb Research tried it.