Ask HN: Is GPT-5 a regression, or is it just me?

[technocratius]

Context: I have been using GPT5 since its release over a month ago, within my Plus subscription. Before this release, I heavily relied on gpt-o3 for most complex tasks, with 4o for simple question. I use it for a mix of scientific literature websearch for e.g. understanding health related topics, the occasional coding assistance and helping me out with *nix sysadmin related tasks. Note that I have not used its API or integration with an IDE.

Based on a month of GPT5 usage, this model feels like primarily like a regression:

1. It's slow: thinking mode can take ages, and sometimes gets downright stuck. It's auto-assessment of whether or not it needs to think feels poorly tuned to most tasks and defaults too easily to going into deep reasoning mode.

2. Hallucinations are in overdrive: I would assess that in 7/10 tasks, hallucinations continuously clutter the responses and warrant corrections and careful monitoring and steering back. It hallucinates list items from your prompt that weren't there, software package functionalities/capabilities and CLI parameters etc. Even thorough prompting with explicit linking to sources, e.g. also wihtin deep research frequently goes of the rails.

3. Not self critical: even in thinking mode, it frequently spews out incorrect stuff, that a blatant "this is not correct, check your answer" can directly correct.

Note: I am not a super advanced prompt engineer, and this above assessment is mainly wrt the previous generation of models. I would expect that with progression of model capabilities, the need for users to apply careful prompt engineering goes down, not up.

I am very curious to hear your experiences.

[incomingpain]

I saw the significant difference in gpt5. If someone were using mostly or just gpt4 before, then it might be a culture shock of difference type situation.

Me who actively uses claude, gemini, perplexity, and a whole gamut of local LLMs.

The personality of the models are different and so when gpt5 came along, it wasnt really a surprise to me.

[lcnPylGDnU4H9OF]

GPT-5 was an upgrade for investors. The primary feature of it is to use a router that will decide between a stronger model and a weaker one for a given query. The goal is to reduce operating costs without regard to improving the user experience, while they market it as "new and improved".

[theturtlemoves]

"Ahhh, you're right, now it's clear, that explains it....."

Yeah, you're not alone. I've even been getting responses with contradictions within the same sentence. ("X won't work therefore you should use X")

[patrakov]

You are not alone.

Another pet peeve is that it, when asked to provide several possible solutions, sometimes generates two that are identical but with different explanations.

[technocratius]

Ah yes, I've had similar experiences actually. Also the variant where I ask it to provide an alternate solution/answer to the one it gave, where it than proceeds to basically regurgitate its previous answer with slight stylistic (i.e. maintaining content parity) modifications.

[android521]

Hallucinations are definitely up at least 5X compared with gpt4 from my personal experience.

[NoahZuniga]

It's just you