Is GPT-5 Getting Dumber? How to Improve AI Performance

8/12/2025

So, You Think GPT-5 is Getting Dumber? Here’s What’s Really Going On.

Ever had one of those days with GPT-5 where it feels like you're talking to a completely different AI? One day it's a genius, spitting out brilliant code & prose, & the next, it can barely understand a simple request. You're not going crazy. Turns out, there's a LOT more going on under the hood than most people realize, & yeah, sometimes you are talking to a different AI.

Here's the thing, GPT-5 isn't just one single, monolithic model. It's a whole family of models, & there's a system in place that's constantly deciding which one is right for your prompt. This is both pretty cool & the source of a lot of frustration. So, let's break down what's happening, why you sometimes get sent to a "mini" model, & what you can do to get the most out of every single conversation.

The GPT-5 Family: It's Complicated

First off, let's get this out of the way: OpenAI didn't just build one GPT-5. They built a whole suite of them, each with different strengths. Think of it like a team of specialists. There are models like

gpt-5-main

, which is the fast, everyday workhorse, & then there's

gpt-5-thinking

, which is the deep reasoner you want for your toughest problems. On top of that, each of these has a

mini

version, & there are even

nano

pro

versions in the mix for different use cases.

So, how does ChatGPT decide which one to use? With a "router." This is a piece of AI that reads your prompt & tries to figure out how much brainpower is needed to answer it. Simple question? It'll likely send you to

gpt-5-main-mini

. A complex, multi-step problem? That should get routed to

gpt-5-thinking

. The goal is efficiency – not every question needs the full, energy-guzzling power of the top-tier model.

This is where the "dumber" feeling comes in. If you're a free or even a Plus user, you have usage caps. Once you hit those limits, you're more likely to be sent to the

mini

models for the rest of your session. This was especially painful during the initial GPT-5 rollout, when a technical glitch with the router made the experience even worse, leading to a lot of users feeling like the model was a major step down from GPT-4o.

How Do You Know Which Model You're Using? Good Luck With That.

This is the million-dollar question, isn't it? And honestly, the answer is a bit frustrating. For the average user, there's no easy way to tell which specific model you're interacting with at any given moment. The interface doesn't have a little light that switches from "genius mode" to "mini mode."

If you're a developer, you can pop open the developer tools in your browser, inspect the network requests, & find something called the

model_slug

in the data stream. But let's be real, who's going to do that for every single prompt?

You might be tempted to just ask ChatGPT, "Hey, which model are you right now?" And you'll get an answer, but it's not always reliable. The model might tell you it's the full GPT-5, even when you're clearly interacting with a less capable version. It's not trying to lie to you, it just might not have access to that information itself.

The "Secret" Phrase to Unlock More Power

So, if you can't always know which model you're on, can you at least try to influence the router? Turns out, you can. It's not a guaranteed, silver-bullet solution, but it can definitely help.

In a Reddit Q&A, an OpenAI researcher let it slip that you can push the system to use its more powerful reasoning capabilities by including a simple phrase in your prompt: "think hard about this."

It's pretty cool, right? By explicitly asking the model to engage its deeper thinking processes, you're giving the router a strong signal that this isn't a simple, lightweight task. I've been playing around with this, & it genuinely seems to make a difference on more complex prompts. If I'm asking for a detailed analysis or a creative solution, adding "think hard about this" at the end often results in a more nuanced & comprehensive response. It's like telling your friend, "Okay, seriously, I need your full attention for this one."

Making the Mini Models Work for You: A Crash Course in Prompt Engineering

Okay, so you've hit your usage cap, or maybe your query is just getting routed to a smaller model. You're not completely out of luck. This is where your prompt engineering skills REALLY come into play. Working with a smaller model is a bit like driving a manual car instead of an automatic – it requires a bit more effort & skill, but you can still get where you're going.

Here are some techniques that are SUPER effective with smaller models:

1. Few-Shot Prompting: Give it Examples

This is one of the most powerful techniques, hands down. Instead of just telling the model what to do, you show it. You provide a few examples of the kind of output you're looking for.

Let's say you want the model to write short, punchy product descriptions. A zero-shot prompt (no examples) would be:

"Write a product description for a new coffee mug."

A few-shot prompt would look like this:

Prompt: "Write a product description for a new product, following the style of the examples."
Example 1: Product: A self-watering planter Description: "Keep your plants happy, even when you're on vacation. This planter knows when your green friends are thirsty."
Example 2: Product: A weighted blanket Description: "It's like a hug you can sleep in. Say goodbye to restless nights."
Your Request: Product: A smart coffee mug that keeps your drink at the perfect temperature. Description:

See the difference? You're giving the model a clear template to follow. This is a game-changer for getting the tone, style, & format you want, especially from a smaller model that might not pick up on the nuances otherwise.

2. Chain-of-Thought Prompting: Make it Show its Work

Smaller models can sometimes jump to conclusions or miss steps in a logical process. Chain-of-thought prompting forces the model to slow down & think things through step-by-step.

Instead of asking:

"I have 10 apples, I give 3 to John, and I buy 5 more. How many apples do I have?"

Try this:

"I have 10 apples, I give 3 to John, and I buy 5 more. How many apples do I have? Think step by step."

By adding that simple instruction, you're encouraging the model to break down the problem: "Okay, you start with 10. Give away 3, that leaves 7. Add 5 more, that's 12." This dramatically reduces the chance of simple arithmetic or logic errors.

3. Be OBSESSIVELY Clear With Your Instructions

This is good advice for any LLM, but it's CRITICAL for smaller ones. You can't assume the model knows what you mean. You have to spell it out.

Specify the format: "Put the answer in a bulleted list." "Give me the output as a JSON object." "Write the response as a friendly email."
Define your terms: If you're using industry jargon or a term that could be ambiguous, define it in the prompt.
Provide context: Don't just ask for a marketing slogan. Tell the model who the target audience is, what the brand's voice is, & what the key selling points are. The more context you provide, the better the output will be.

When Consistency is Non-Negotiable: The Case for Custom AI

This whole song & dance with model routing & prompt engineering is fine for personal use, but for a business, that kind of inconsistency can be a HUGE problem. Imagine if your customer support chatbot was a genius one minute & completely unhelpful the next. It would be a nightmare for your brand.

This is where platforms like Arsturn come in. Honestly, for any business that's serious about using AI for customer interaction, building your own custom chatbot is the way to go. With Arsturn, you can create a no-code AI chatbot that's trained specifically on YOUR data. This means it's not guessing based on the vast, messy internet – it's an expert on your products, your policies, & your customers' questions.

You get total control over the AI's personality & responses. You're not at the mercy of some mysterious internal router deciding to send your customer to a "mini" model right when they have a complex problem. The AI provides instant, accurate support 24/7, because it's built for one purpose: to be an expert on your business. It's a way to get all the benefits of AI automation without the unpredictability of a general-purpose model. For things like lead generation, customer engagement, & website optimization, having a consistent, reliable AI is a TOTAL game-changer.

The Bigger Picture: An Evolving Relationship

The good news is that OpenAI is listening to user feedback. After the initial rocky rollout of GPT-5, they've acknowledged the issues, doubled the rate limits for Plus users, & have even said they're considering giving users the option to stick with older models like GPT-4o if they prefer. This is a sign that we're in a collaborative, if sometimes bumpy, process of figuring out how to best work with these incredibly powerful tools.

So, the next time GPT-5 feels a little...off...don't just throw your hands up in frustration. Remember what's happening behind the scenes. Try to nudge the router with a "think hard" command. If that doesn't work, roll up your sleeves & put your prompt engineering skills to the test. And if you're a business looking for a more stable & reliable solution, seriously, look into building your own custom AI with a platform like Arsturn. It's the best way to ensure your AI is always in "genius mode" when your customers need it.

Hope this was helpful! Let me know what you think, & if you have any other tips for getting the most out of GPT-5, I'd love to hear them.