Sam Altman promised "PhD-level intelligence." Users got a model that thinks Oregon is called "Onegon" and Joe Biden is still president.
OpenAI's GPT-5, billed as the company's most advanced AI yet, has sparked widespread disappointment and ridicule since its Thursday launch, with users reporting basic errors that have reignited debates about whether the AI industry's scaling approach has hit a wall.
The backlash was swift and brutal. Social media lit up with screenshots of GPT-5's bizarre errors, like maps labeling Oklahoma as "Gelahbrin," basic math problems like "5.9 = x + 5.11" solved incorrectly, and timelines featuring fictional presidents like "Willian H. Brusen."
"The overwhelming consensus on GPT-5 from both X $TWTR and the Reddit $RDDT AMA are overwhelmingly negative," noted the AI Leaks and News account, while an informal poll found most users rating the model as simply "Kinda mid."
Yet the reception wasn't uniformly negative. OpenAI reported that API traffic had doubled within 24 hours of launch, and some early users praised GPT-5's coding abilities and creative output. Box CEO Aaron Levie highlighted major improvements in extracting data from complex legal documents. Wharton professor Ethan Mollick noted the AI's ability to anticipate user needs and deliver comprehensive results beyond basic requests. Developer Simon Willison wrote that GPT-5 is "my new favorite model."
However, those positive reviews were overshadowed by the market's harsh verdict. On Polymarket, a prediction platform where traders bet on future events, OpenAI's odds of having the best AI model by month's end collapsed from 75% to 14% in a single hour Thursday night.
By Friday, a contrite Sam Altman was doing damage control on Reddit after users posted comments like "GPT-5 is wearing the skin of my dead friend" and started a Change.org petition. He promised to restore access to the previous version, GPT-4o — which OpenAI did by Friday evening — and admitted that a broken "autoswitcher" between GPT-5's different modes had made the model "seem way dumber" than intended. The mea culpa marked a stunning reversal for a CEO who just 24 hours earlier had declared his latest creation "clearly a model that is generally intelligent."
The debacle raises fundamental questions about whether the AI industry's core strategy of building ever-larger models has hit a wall. Critics have long argued that simply scaling up models won't lead to artificial general intelligence, and GPT-5's stumbles seem to validate those concerns.
"My work here is truly done," wrote Gary Marcus, an AI researcher and a longtime critic of current LLM development approaches. "Nobody with intellectual integrity can still believe that pure scaling will get us to AGI." The problems coming from what was billed as a doctorate-level tool suggest fundamental limitations in how large language models process information, not just teething troubles that can be fixed with patches.
The timing couldn't be worse for OpenAI. Alibaba’s AI group announced its Qwen 3 model with 1 million token context, allowing for conversations nearly four times what GPT-5 offers. Meanwhile, Anthropic's Claude Opus 4.1 is keeping pace in coding benchmarks, and Gemini, after a late start, is gaining ground. What was supposed to cement OpenAI's dominance instead handed rivals a gift.
The stakes are enormous for a company that just raised $8.3 billion at a $300 billion valuation and is burning through cash on compute costs. OpenAI's annual revenue hit $13 billion, but the company remains unprofitable, making every misstep costly. The botched launch also comes as competitors circle, with Meta $META reportedly offering billion-dollar packages to poach OpenAI talent.
OpenAI plans to roll out new productivity features next week, including Gmail and Google $GOOGL Calendar integration, and has historically smoothed out launch issues over time. But for a model that the company had worked on since late 2023, the reception is probably not what OpenAI was hoping for. In January, Altman wrote that "we are now confident we know how to build AGI" — a confidence that seems premature given a model that can't correctly label Oregon on a map.