OpenAI Chief Executive Sam Altman had teased GPT-5 for months. He claimed it was an advanced model – smarter, faster, and capable of thinking at a “PhD level.” The company positioned that launch as a major step forward for ChatGPT — one meant to lead to intelligence improvements in coding, reasoning, and accuracy.
But early reactions paint a muddier picture. Developers praised the model for understanding complex prompts and outputting well-structured code. Early tester Simon Willison described GPT-5 as “competent” and “occasionally impressive”, but not a huge leap from GPT-4. Others were less impressed.
Several posts on social media quickly turned to complaints about numerous factual errors, weak math skills, and — in a few cases — even basic spelling mistakes. Noah Giansiracusa, a math professor at Bentley University, called the release “underwhelming,” noting that the updates in question felt “more marginal than I would’ve hoped.”
Part of the confusion was due to the architecture of the model. GPT-5 would include an ‘autoswitcher’ for the various model sizes, depending on its task. This saves processing and means you are not always touching the full GPT-5, which back-doored many people. Upon answering a question incorrectly with the system, the agent it was working with instructed it to “think harder” about how many “b” letters are in blueberry. After that feedback, it got the answer right when queried.
Users push back—and OpenAI responds
The frustration could spill onto Reddit and X by Friday. And while some users hated the fact that they wouldn’t know who or even which model the text came from, many felt like hell was just an inference, and GPT5 replaced old favorites they trusted. A few said the quality has been affected, writing is not as good as GPT-4.5, which you agreed should be, and some creative and technical things felt worse.
This chat led the CEO of OpenAI to participate in a Reddit “Ask Me Anything” that would address the backlash. He said a bug in the autoswitcher meant GPT-5 ran below its full potential on day one. He promised to be more transparent in future model transitions: “from today on, GPT-5 should look increasingly smarter. OpenAI also promised to redeploy earlier models such as GPT-4o for those who expressed a preference, an effort to appease the loudest critics.
That it needed to act this quickly highlights just how big the stakes were. Today, OpenAI is in a breakneck race with competitors such as Anthropic, Google DeepMind, and even Elon Musk’s xAI. With all this competition, a little slip-up, and soon your competitor is taking a step ahead of you.
Competition intensifies in the AI race
Even with such differences, the reception to GPT-5 was lukewarm, but that did not stop it from being funneled up to the top spot on LMArena, a user-driven leaderboard for training models. While it wasn’t what we would consider today as programmable, it was extremely fast and expressive, supporting large, complex queries. However, other benchmarks showed a different, more challenging problem landscape. Grok 4 from Musk’s xAI scored so high on ARC-AGI-2 (a set of tests for advanced reasoning) that it presented a significant challenge to OpenAI’s supremacy.
Some industry analysts speculate it will be at least a day before the effects of GPT-5 are fully understood, similar to previous breakthroughs by the firm, such as GPT-3. GPT-5 and GPT-4 received significant backlash before they saw improvements in updates and other use cases. The model has the potential to succeed (or fail) depending on how useful it is for people’s everyday workflow, and there are nearly 700 million weekly ChatGPT users around the globe.
In the words of Wharton professor Ethan Mollick: “GPT-5 just does stuff, sometimes amazing things, sometimes puzzling things, entirely on its own. That is what renders it so intriguing. It’s still a technocratic, slightly creepy approach to calling things like this part of the public health sphere “interesting”
Want your project in front of crypto’s top minds? Feature it in our next industry report, where data meets impact.
