Simple bench is a website for testing AI models on questions that are made in such a way to be easy and simple for humans, even if they are not any kind of experts - and yet difficult and tricky for AIs. Typically they involve using a lot of common sense, having feeling for priorities, and deep understanding of the world.
So far the best score achieved on this benchmark is 62.4% by Gemini 2.5 Pro.
I'll resolve it as soon as the results for GPT 5 on this test come out.
If by the end of 2026 the results don't come out, I'll resolve it as NO.
The question is concerning only basic model known as GPT-5 and all more advanced versions (for example if there's some "reasoning" model) that are released at the same time as the main model itself.
Later updates (like GPT 5.5 or GPT 5o) don't count, as they will be treated as entirely different models.
1,000