Five days ago, I published “Five Hours with GPT-5,” calling OpenAI’s latest model “impressive.” I was wrong. After living with GPT-5 for more than a week, my user experience tells a different story than the initial benchmarks. I’m not alone.
The Prose Problem
GPT-5 has developed an obsession with bullet points and lists that borders on pathological. Ask for a business memo, get bullets. Request analysis, receive numbered lists. Even when you explicitly demand “no bullets, no lists, no em dashes,” it sneaks them back in.
I spent twenty minutes yesterday trying to get some copy to match the style of a context profile I provided. It took twenty minutes because I was absolutely sure the next iteration would be what I needed. GPT-4o would have nailed this on the first try.
My breaking point was a simple flowchart edit. Eight attempts to add one step to an existing workflow. Eight failures. I gave up and created it in 30 seconds using OpenAI’s image generator. Their image model understood the task better than their “fully multimodal,” “PhD-level” flagship model.
The Router Ruse
Here’s what OpenAI didn’t mention: the router broke on launch day, making GPT-5 seem “way dumber,” according to CEO Sam Altman. The “unified system” was cost-cutting disguised as innovation.
The router was supposed to intelligently route queries between fast and deep-thinking models. In practice, users waited two minutes watching elaborate reasoning animations, then received output worse than GPT-4o’s instant responses. Meanwhile, the system served most queries to cheaper models while charging full price.
The Walkback Parade
User backlash was swift. Nearly 5,000 users flocked to Reddit with complaints like “GPT-5 is horrible” and “feels like a downgrade.” OpenAI’s response was a series of embarrassing retreats:
Day 1: Launch GPT-5 as the unified future, deprecate all previous models
Day 2: Restore the model picker with Auto, Fast, and Thinking options
Day 3: Bring back GPT-4o after user protests
Day 4: Increase rate limits from 200 to 3,000 messages per week
Day 5: Altman admits deprecating models users depended on “was a mistake”
The Real Strategy
Users figured it out quickly. “They removed all their expensive, capable models and replaced them with an auto-router that defaults to cost optimization,” wrote one Reddit user. “That sounds bad, so they wrap it up as GPT-5 and proclaim it’s incredible.”
OpenAI eliminated model choice, limited free users to 10 messages per hour, unchanged context windows despite claiming major advances. If OpenAI needed to prove profitability to justify their $500 billion valuation, calling cost-cutting an “upgrade” probably wasn’t the way to do it.
Reading the Tea Leaves
OpenAI’s financial pressures are becoming clearer. Generating high-quality AI responses costs more than they are charging for them. Nick Turley, OpenAI’s VP of ChatGPT, recently said that OpenAI wouldn’t rule out advertising in the future.
So, are we witnessing a temporary misstep from a company racing to maintain market leadership, or the emergence of a business model that prioritizes cost optimization over user experience?
More fundamentally, can the AI industry sustain its current trajectory when the economics don’t add up? And if companies like OpenAI start choosing efficiency over effectiveness as their north star, what does it mean for enterprises betting their digital transformation strategies on these platforms? (Importantly, OpenAI has maintained all of its APIs. No enterprise applications were impacted by the GPT-5 misstep.)
In the end, my five days with GPT-5 reinforced that user experience matters as much as benchmark scores. Hopefully, OpenAI learned the same lesson.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.