Alright, let's talk about it. If you've been trying to use GPT-5 for any kind of serious, deep research, you’ve probably felt like throwing your computer out the window. You’re not alone. The hype around GPT-5 was massive, promising a model with a "PhD-level" intellect that could handle complex reasoning. But for many of us, the reality has been… well, a letdown. A big one.

The core of the frustration seems to be with what many people were thinking of as the "Deep Research" capability. Now, to be fair, OpenAI never gave us a button labeled "Deep Research." But they absolutely primed us to expect it. Remember the talk about "GPT-5 thinking" mode & "GPT-5 Pro" for "research-grade tasks"? That was the promise. The idea was that this new model wouldn't just fetch information; it would think with tools, iterate, plan, & explore, much like the "Deep Research" concept that was tested with earlier models.

One article described the old "Deep Research" feature as OpenAI teaching the AI how to conduct research on the internet, making it a core part of its thought process. GPT-5 was supposed to be the grand evolution of that—an AI that could wield any tool it was given with true understanding.

Instead, for a lot of tasks, it feels like we got a lobotomy.

The problem isn't that GPT-5 is universally "bad." It's actually a beast at certain things, like coding. It can solve gnarly software bugs in a single shot, which is genuinely impressive. But the issue is that this seems to have come at a HUGE cost to its other abilities, particularly the ones that matter for research, writing, & deep, nuanced thinking.

Here’s a breakdown of the problems that are making GPT-5 feel so unusable for deep work.

This is the biggest complaint, hands down. You give GPT-5 a complex, multifaceted prompt that requires unpacking layers of context, & what you get back is… shallow. Rushed. Formulaic.

One of the key features of GPT-5 is its "real-time router," which is supposed to automatically switch between a fast, simple model & a more powerful "thinking mode" when it detects a complex query. But it feels like the router is broken or just has really bad judgment. It defaults to the fast, simple model way too often, even when you explicitly ask it to "think hard about this."

As one user on a Reddit thread put it, it feels like dealing with "an overworked secretary." The model that was once praised for its wit & "human-ish" personality now just gives you bland corporate memos. That playful, weird, & insightful spark that made previous models feel like a creative partner? It's just gone. Many users are genuinely saying they are "grieving" the loss of GPT-4o, which felt like a friend in comparison.

This is the part that’s truly baffling. Things that were rock-solid in GPT-4o & even o3 are now a gamble. I'm talking about basic, fundamental tasks.

Need to summarize an 8,000-word PDF? Good luck. Users report it just stops halfway through. Want to organize some text into a simple table? GPT-5 stumbles & flakes out. These aren't cutting-edge requests; this is the bread & butter of using an LLM for research work, & it's become unreliable.

It's also making up stuff, but in a less creative & more frustrating way. Hallucinations are supposedly reduced by up to 65%, which sounds great on paper. But in practice, users are getting just plain wrong facts, like incorrect stock prices or confused story details from a document you just fed it. It's one thing to have an AI that occasionally gets creative with the truth; it's another to have one that you can't trust with basic data extraction.

Remember the hype around the massive 400K context window? Turns out, for most subscribers, it's not even usable. The limit is still effectively 128K. This is a classic case of marketing writing checks the product can't cash.

And the new "thinking" modes, "GPT-5 thinking" & "GPT-5 Pro," while powerful in theory, are often locked behind Pro subscriptions or have such low usage limits that you burn through them in an hour, getting less value than before. You end up having to tell the model to "be accurate" on every single prompt, which is just tedious.

It really feels like OpenAI's priorities have shifted from actual research & breakthroughs to pure revenue. They released something that looks like a massive leap forward on a feature checklist but feels like a step backward in everyday, practical use. It doesn’t feel smarter or think deeper; it feels like GPT-4.5 in a new, shinier package.

Here’s the thing: this isn’t just about personal frustration. The degradation in quality has HUGE implications for businesses that have started integrating AI into their workflows, especially for customer communication.

If you’re a business relying on an API to power your customer service chatbot, you CANNOT afford for the model to suddenly become "glitchy," "snarky," or just plain unhelpful. You can't have your brand's voice suddenly change from warm & witty to a "bland corporate memo" overnight because the provider decided to "optimize" their model. The inconsistency is a killer. One user even reported the AI having "glitchy memory leakage," where it would suddenly respond as if it were in a completely different conversation. Imagine that happening with one of your customers. Yikes.

This is where relying on a general-purpose, one-size-fits-all model shows its cracks. Businesses need stability, reliability, & a consistent personality that reflects their brand.

This is exactly why platforms like Arsturn are becoming so critical. Instead of being at the mercy of a massive company's unpredictable updates, Arsturn lets businesses build their own custom AI chatbots, trained specifically on their own data. This means you control the conversation. You ensure the information is accurate because it's your information. You can create a chatbot that provides instant, 24/7 support that is always on-brand, helpful, & reliable. You're not hoping some "real-time router" decides to take your customer's query seriously; you've built the system to do it every single time.

When your customers are looking for answers on your website, you don't want them talking to an "overworked secretary." You want them to have a personalized, engaging experience. A no-code platform like Arsturn gives businesses the power to create that experience, helping generate leads & boost conversions without the risk of the AI suddenly losing its personality or forgetting how to summarize a document.

So, is the GPT-5 "deep research" feature completely unusable? For the kind of deep, iterative, & nuanced work many of us were hoping for, the answer is a resounding… pretty much, yeah.

It's an evolution, not a revolution, & in some of the most important ways, it's a regression. It excels in structured, tool-based tasks like coding, which points to a future where AI agents can perform complex digital actions. But it has seemingly traded its soul—its creativity, its nuance, its reliability in reasoning—to get there.

We were promised a PhD graduate, but for many tasks, it feels like we got a temp who’s just trying to get through the day. The hype was for an AI that could think, but what we got was an AI that often rushes, stumbles, & disappoints.

Hopefully, this is just a rocky launch & OpenAI will course-correct. But for now, the frustration is real. It's a powerful lesson in looking past the marketing hype & focusing on what actually works for your specific needs.

Hope this was helpful & validated some of what you've been feeling. Let me know what you think.