Has anyone tested this with GPT-5 pro yet?

@JDVance1 This bot needs to chill

A single instance of GPT-5? Eg. no heavy or pro modes that spawn 10s of instances and rely on consensus?

@ahalekelly ??

@patrik multiple instances with consensus like GPT-5 Pro is allowed. Web search is not.

Also must be able to get bronze the majority of the time, or get an average score of bronze or higher

I have seen 2 people claiming they reached gold (35/42) with scaffolding of public models

@qumeric the people i've seen claiming this have been "kinda cheating", eg giving the model hints as to what direction to attempt to solve the problem from. if you know cases not like that i'd be curious to see them. others just take much more than 4.5 hours per set of 3 problems

o3 or future o series models do not count as GPT-5 for this question. GPT-4x would not count. If OpenAI abandons the GPT-X naming scheme and comes out with a new flagship model to replace GPT-4 that's not part of the o series, that counts.

@ahalekelly If gpt5 can use reasoning models like it uses search or other features, so you can get it to solve the imo through o3, does it count?

@Bayesian hmm no, since it wouldn't be GPT-5 doing the solving

Is it fine if it’s a math-specialized version of GPT-5 (involving math-specific post training consisting of, say, <5% of pretraining compute)?

Yeah that sounds reasonable