GPT-5: The Launch That Changed Everything (And Confused Everyone)

The anticipation for GPT-5 has been building for months.

And in true OpenAI fashion, they released GPT-5 just two days after launching their open source models… Talk about dominating the news cycle.

When GPT-5 finally arrived, it brought something unprecedented: advanced reasoning capabilities for everyone, including free users. But it also delivered one of the most chaotic AI launches I've witnessed in years.

So today, I’m breaking down the whole story:

• Why GPT-5 represents a new era in AI accessibility

• The surprising backlash and what it revealed about user behavior

• My early testing experience as a developer (spoiler: the model is impressive, but the tooling needs work)

—@dharmesh

Why GPT-5 Marks a New Era

In my opinion, the biggest piece of the GPT-5 launch wasn't the technical capabilities, it was the democratization of advanced AI reasoning.

While every AI enthusiast is screaming for "AGI," the reality is that the majority of the world is still just getting used to living with ChatGPT and these new AI systems in their daily lives.

For the first time, OpenAI made their most sophisticated reasoning model available to everyone, including free users. This is a big shift.

Sam Altman shared a shocking statistic: <1% of free users were using OpenAI’s smartest reasoning models. 700M weekly active users, and only a tiny sliver of that was experiencing the AI's true capabilities because of complexity barriers.

With GPT-5, those barriers are disappearing. All users get access to the world’s smartest reasoning model.

The rollout includes three variants:

• GPT-5: The standard model

• GPT-5 Pro: Enhanced capabilities for complex tasks

• GPT-5 Mini: Optimized for speed and efficiency

Most cleverly, GPT-5 includes an automatic router that switches reasoning on and off based on task complexity. You don't have to think about whether you need "reasoning mode"--the model figures it out for you.

The performance improvements are across the board: state-of-the-art results in coding, writing, mathematics, and health benchmarks.

But based on my own "vibe tests" over the past week, what impressed me most is how much less these models hallucinate and how honestly they communicate when they can't handle a task.

The Backlash (And What It Revealed)

If you try GPT-5 now, it works beautifully. But the launch was a different story.

What should have been a celebration quickly turned into chaos that revealed just how dependent we've all become on these AI systems.

The problems started immediately:

• The automatic switcher crashed on launch day, making GPT-5 appear significantly less capable than it actually was

• Users encountered touch mode glitches and frustratingly low rate limits

• Disappointment that GPT-5 wasn't the mythical "AGI" everyone was expecting

But the real shock was the user revolt. Reddit exploded with complaints demanding the return of GPT-4o. OpenAI had deprecated previous models to streamline the experience, but users weren't having it.

OpenAI has now brought back GPT-4o through the Legacy models toggle, but I think the intensity of the reaction revealed something profound: people had developed genuine emotional attachments to specific AI model behaviors.

But here’s what confused me the most: the "fix" for most GPT-5 complaints was embarrassingly simple. If you wanted the model to think more deeply, you just needed to add "think deeply" to the end of your prompt. This essentially forced the router to use its highest-powered reasoning mode.

This whole situation made me curious about how many ChatGPT features people actually know about.

So I did what any data-obsessed person would do and posted a LinkedIn poll about Custom Instructions usage:

The results were eye-opening: 30% of respondents didn't even know what Custom Instructions are. Keep in mind, my LinkedIn audience skews heavily AI-forward since that's all I talk about. This suggests there are probably dozens of valuable ChatGPT features that most users have never explored.

Which brings me to a question for you: Are there any tactical ChatGPT features you'd like me to deep-dive into? I'm thinking of doing a "hidden features" series.

Early Developer Testing: GPT-5 + Cursor

I've spent the past week testing GPT-5 extensively, particularly with Cursor (the AI-powered code editor), and I have some thoughts.

The Good News: GPT-5 as a model is really good. It feels within range of Anthropic's Claude Sonnet and Opus models, which until now were my go-to choices for complex coding tasks. The reasoning capabilities shine when working through multi-step programming problems.

The Reality Check: Despite GPT-5's capabilities, the overall experience of using Claude Code is still better for most development workflows. This isn't necessarily about the underlying model--it's the implementation and user experience design.

The Opportunity: Since Cursor is the IDE (being a fork of VS Code), they have a unique opportunity to create deeper integration between the AI and the development environment.

Instead of dumping reasoning and work logs into a chat panel, imagine if it:

• Showed the current file being worked on directly in the IDE as modifications happen

• Auto-scrolled to relevant code sections being written or modified

• Visually indicated which files are active in the file explorer

• Used the IDE's built-in search instead of grep commands (not because it's better, but because it mirrors how human developers actually work)

If they wanted to go really deep, they could add "visual annotations" to modified files--like comments in Google Docs. That would be genuine UI innovation.

Right now, the IDE feels like it's mainly there so I can make direct changes to generated code. But there's potential for much deeper integration where the AI feels like it's working within your development environment rather than alongside it.

The potential is there. Someone just needs to fully realize it.

What This All Means

The GPT-5 launch was messy, but historic: the moment advanced AI reasoning became accessible to everyone.

The backlash taught us that AI adoption isn't just about smarter models anymore--it's about change management and understanding that people develop deep preferences for how their tools behave.

From a developer perspective, models are getting dramatically better, but tooling is still catching up. Companies that figure out seamless AI-workflow integration will have significant advantages.

Most importantly, when 700 million people suddenly have access to advanced reasoning capabilities, the innovation that follows will be unprecedented.

It’s time to build!

—Dharmesh (@dharmesh)