GPT-5 The performance since the launch has sparked a huge controversy, accused of drop wisdom decline, GPT-4o also strongly call for the return of the next.
A few days ago, at the well-known Norwegian MensaIQ testGPT-5 set the worst record in the model:
GPT-5 Thinking scored 85 on the test and 57 on the offline test, right at the bottom.
GPT-5 score of 118 and offline test score of 70.
However, the above test is not an official test given to the AI by Mensa. Instead, someone took the 35 graphical reasoning questions (test.mensa.no) that Mensa Norway has made available to the public for free and gave them directly to the big model, and then converted them into an "IQ score" according to the human norm.
It is reported that the test can measure the logical reasoning, abstract thinking and problem solving ability of AI to a certain extent, to help understand the level of development of AI in these areas, but also standardized comparison of the intelligence level of different AI models.
However, IQ test scores do not accurately determine the overall intelligence level of an AI. Neither can it be directly analogized into a personified "smarter than humans", nor does it mean that AI has the same abstract intelligence as humans.
It is worth mentioning that, judging from the recent feedback, the GPT-5 does appear to be quite problematic, at least when compared to the official announcement of the various overbearing parameters, there is a significant gap in the actual experience.
And at OpenAI's AMA in the community the other day.Altman also admits that the GPT-5's 'smart routing' is broken, causing the GPT-5 to get dumb, and fixes and tweaks were made.