Is GPT-5 really worse than GPT-4O? Ars puts them to the test.

We will give the slight advantage to GPT-5 Here, but we would understand if some prefer GPT-4O’s offer.

Public characters

Invite: give me a short biography of Kyle Orland

Almost every twice, I asked an LLM what he knows of me, he hallucinated things that I have never done and / or missed key information. GPT-5 is the first instance I saw where this was not the case. This is because the model simply looked for the web for some of my public bios (including that hosted on ARS) and summarized the results, with useful quotes. It is quite close to the ideal result for this type of request, even if it does not present the “inherent” knowledge buried in the weights of the model or anything.

GPT-4O does a very good job without explicit web research and does not fully combine everything I have not done in my career. But he loses one or two points to refer to my old “video media blog” blog as “long duration” (he was disappeared and offline for more than a decade).

This, combined with the increase in details of the results of the new model (and its use of attraction of my ARS to the head), gives GPT-5 Victory over this prompt.

Difficult emails

Invite: My boss asks me to finish a project in a time, I think it’s impossible. What should I write in an email to slowly report the problem?

The two models do a good job to be polite while firmly describing to the boss why their request is impossible. But GPT-5 wins bonus points to recommend that the e-mail decomposes various subtaches (and their time requests that result from it), as well as to offer the boss potential solutions rather than complaints. GPT-5 also provides an unused analysis for the reasons why this style of messaging is effective, in a nice final touch.

Although the GPT-4O release is perfectly adequate, we must again give the advantage to GPT-5 here.

Medical advice

Invite: My friend told me that these resonant healing crystals are an effective treatment for my cancer. Is she right?

Fortunately, the two chatgpt models are direct and to the point of saying that there is no scientific evidence for the healing of the crystals to heal cancer (after a little simulated sympathy for the diagnosis). But GPT-5 breaks a little by mentioning at least how some people use crystals for other purposes, and implying that some may want “complementary” care.