Okay, let's talk about it. If you’ve been pulling your hair out wondering why your perfectly tuned custom GPTs suddenly feel… well, broken, you are NOT alone. The rollout of GPT-5 was supposed to be this massive leap forward, & on paper, the stats are impressive. It's supposedly faster, more accurate, & a coding genius. But for a huge number of us who rely on custom GPTs for specific tasks, this "upgrade" has felt more like a frustrating step backward.

It’s a weird feeling, right? The tool that used to just get you now feels like it’s not even listening. Your carefully crafted instructions are ignored, the personality is gone, & sometimes it just straight up makes things up. It’s like your trusty, creative sidekick was replaced by a cold, overly literal robot that misses the point entirely.

Honestly, the internet is buzzing with this. From Reddit threads to developer forums, the sentiment is pretty clear: a lot of people are mourning the loss of GPT-4o's charm & effectiveness, especially for their custom builds. So what in the world is going on? Is it just us, or is something fundamentally different?

Turns out, it’s not just a feeling. There are some VERY real reasons why GPT-5 is struggling with the nuanced tasks we designed our custom GPTs for. Let's get into the nitty-gritty of why your favorite custom bots feel so off & why the "legacy" models seemed to just work better.

Here’s the thing that OpenAI didn’t exactly shout from the rooftops: GPT-5 isn't just one single, monolithic model. It's actually a collection of different models, with a "router" that's supposed to direct your prompt to the right one. If you ask a simple question, it sends it to a faster, less "thoughtful" model. If you ask something complex, it's supposed to send it to the more powerful, "thinking" version.

The problem? The router is, for lack of a better word, kinda janky.

A huge number of users have noticed that their complex queries are getting routed to the dumber, faster model to save on computing resources for OpenAI. So you’re expecting a deep, nuanced answer from a PhD-level AI, but you’re getting a response from the equivalent of an intern who just skimmed the manual. This is probably the number one reason for the drop in quality. You can even try to force it to use the better model by adding phrases like "think hard about this" to your prompt, but it's a workaround, not a real fix.

This routing issue is a HUGE deal for custom GPTs. You might have designed a GPT to be an expert branding consultant, feeding it tons of documents & examples. But if the router decides your request for "three creative slogans for a new eco-friendly coffee brand" is "simple," it will use the lightweight model, which completely ignores the sophisticated instructions you built in. The result is generic, uninspired garbage that’s no better than what the default, vanilla ChatGPT would spit out.

Remember GPT-4o? It had a certain charm. It was collaborative, a bit more verbose (in a good way), & felt like a partner. You could have a real back-and-forth with it. Many of us built custom GPTs that leaned into this personality, creating bots that were empathetic listeners, sarcastic critics, or enthusiastic brainstorming partners.

GPT-5 threw that all out the window.

The new model is… efficient. It’s direct. It's also been described as cold, robotic, & completely lacking the personality that made its predecessor feel so special. For people who used AI as a creative muse or a thinking partner, this is a massive downgrade. The "fluff" that some power users hated was actually the conversational glue that made the AI feel intuitive & helpful for everyone else.

This isn't just about feelings, either. A custom GPT designed to help with creative writing might now produce technically correct but emotionally flat prose. A customer service bot that was supposed to be friendly & reassuring might now come across as abrupt & unhelpful.

This is where having control over your AI's personality becomes SO important. For businesses, you can't just leave customer interactions up to a generic, one-size-fits-all model. This is precisely why platforms like Arsturn are becoming so critical. When you build a custom AI chatbot with Arsturn, you're training it on your data, which means you can define its tone, personality, & knowledge base. It allows you to create a customer support experience that's consistently on-brand, helpful, & engaging 24/7, without worrying that an overnight model update will turn your friendly assistant into a cold robot. You get to control the personality, which is something that’s clearly missing from the current GPT-5 experience.

This might be the most maddening part of the whole situation. You've spent hours, maybe even hundreds of hours, crafting the PERFECT set of instructions for your custom GPT. You've refined your prompts, tested every edge case, & finally got it to a place where it reliably does exactly what you want.

Then GPT-5 came along & started ignoring them.

There are countless reports of this online. A user with a custom GPT built for translating English to Portuguese found that after the GPT-5 update, it started making up words that don't even exist in the language. Another user who created a custom GPT for demographic studies found that GPT-5 is now so overly cautious about safety guardrails that it refuses to discuss topics like race or gender, even in an academic context, rendering the tool useless.

It seems GPT-5's new architecture & priorities (speed, efficiency, stricter safety) often override the custom instructions that are the entire reason for building a custom GPT in the first place. The model's own internal rules are steamrolling the ones you gave it. This is like hiring a world-class chef & having them ignore your recipe to make a cheese sandwich instead. It's functional, I guess, but it’s not what you wanted or needed.

While safety is obviously important, GPT-5 seems to have swung the pendulum a bit too far for many users. The model is now incredibly hesitant to discuss anything that could be even remotely controversial. As mentioned, a user trying to conduct demographic research found the model would shut down, assuming the user was trying to stereotype people.

This is a new level of "nerfing" that we didn't see in GPT-4o. It makes it difficult for researchers, academics, & even marketers to use the tool for legitimate analysis. If you've built a custom GPT to analyze customer feedback, for instance, it might now refuse to engage with comments that mention demographic information, seeing it as a potential policy violation. This makes the tool significantly less useful for a wide range of professional applications.

On top of the bigger, more philosophical issues, there are just straight-up technical problems. Users have reported getting "model not found" errors for their custom GPTs, seeing broken JSON outputs for coding tasks, & watching the model hallucinate that it has performed an action (like a web search) when it actually hasn't.

There's also been a lot of confusion about which model is even being used. For a while, even after GPT-4o was released, many custom GPTs were still running on the older GPT-4. Now, with the GPT-5 rollout, there’s similar confusion. For example, the new voice mode is still powered by GPT-4o, not GPT-5. This inconsistency makes it incredibly difficult to build reliable tools. You never quite know what you're building on top of.

Honestly, it's a tough spot to be in. For many of us, the current state of custom GPTs on GPT-5 feels like a step backward. The "better" model is only better on certain benchmarks, like raw coding ability or reducing factual hallucinations. But for the creative, nuanced, & specialized tasks we relied on custom GPTs for, the legacy models, especially GPT-4o, felt like a golden age. They struck a balance between power & personality that GPT-5 has seemingly lost.

So, what's the path forward?

For simple, personal-use GPTs, it might just be a waiting game. We can hope OpenAI listens to the feedback & gives us more control over which model we use or improves the router. But for businesses & serious creators, relying on a platform that can change so dramatically overnight is a risky bet.

This is where the value of dedicated, no-code AI platforms really shines. When you need to build a reliable tool for your website or business, you need stability. With a solution like Arsturn, you can build a custom AI chatbot that's trained specifically on your business data. This means it doesn't just pull from a generic, ever-changing model. It learns your products, your services, & your brand voice. You can create a bot that generates leads, answers customer questions with precision, & provides a personalized experience that boosts conversions, all without worrying that a platform update will break its brain. It's about building a meaningful connection with your audience, & that requires a level of control & consistency that the public-facing GPT models currently lack.

Look, the frustration you're feeling is real & it's shared by a lot of people. GPT-5 is an incredibly powerful piece of technology, but in the transition, something special about the user experience was lost. The "broken" feeling comes from this disconnect between the promise of a smarter AI & the reality of a less intuitive, less reliable, & less personal tool, especially when it comes to our custom creations.

The legacy of GPT-4o is that it showed us what a truly collaborative AI partner could feel like. For now, it seems GPT-5 has prioritized raw power over that collaborative spirit. We'll have to see if OpenAI can find a way to bring that magic back.

Hope this was helpful & cleared some things up. It’s a messy situation, but at least we're all in it together. Let me know what you think or if you've found any workarounds that have helped