OpenAI admits GPT-5 rollout issues, reinstates GPT-4o for users

OpenAI co-founder and CEO Sam Altman has publicly acknowledged significant challenges following the recent rollout of GPT-5, the company’s new flagship large language model (LLM). Heralded as OpenAI’s most powerful and capable model to date, its launch has been marred by a range of issues, including faulty model switching, inconsistent performance, and widespread user confusion. These problems have prompted OpenAI to partially reverse some platform changes, notably reinstating user access to earlier models like GPT-4o.

In a candid exchange with users on Reddit and X (formerly Twitter), Altman admitted the launch was “a little more bumpy than we hoped for.” He attributed errors in performance charts shown during OpenAI’s GPT-5 livestream to human error, noting that “people were working late and were very tired.” While the accompanying blog post and system card were accurate, these missteps further complicated a launch already facing scrutiny from early adopters and developers. OpenAI has since moved to double GPT-5 rate limits for ChatGPT Plus users and will continue to allow Plus users to opt for GPT-4o, monitoring usage as it considers how long to support legacy models.

A primary cause of the disruption, according to Altman, was a malfunction in OpenAI’s new automatic “router.” This system is designed to assign user prompts to one of four GPT-5 variants—regular, mini, nano, and pro—with an optional “thinking” mode for more complex reasoning tasks. Altman revealed that a crucial component of this system, the “autoswitcher,” was “out of commission for a chunk of the day,” leading GPT-5 to appear “way dumber” than intended. In response, OpenAI plans to adjust the model decision boundary for better performance and will enhance transparency by indicating which model is responding to a given query. A user interface update is also in the pipeline to facilitate manual triggering of the thinking mode.

Despite OpenAI’s internal benchmarks suggesting GPT-5 leads the pack of large language models, real-world user experiences have painted a different picture. Since the launch, numerous users have shared examples of GPT-5 making fundamental errors in mathematics, logic, and coding tasks. For instance, data scientist Colin Fraser posted screenshots illustrating GPT-5 incorrectly determining that 8.888 repeating equals 9, while another user highlighted its failure on a simple algebra problem. Developers, too, have expressed disappointment, with some reporting that GPT-5 performed worse than rival AI lab Anthropic’s new Claude Opus 4.1 in “one-shot” programming tasks, which require completing a task accurately from a single prompt. Furthermore, security firm SPLX identified that GPT-5 remains vulnerable to prompt injection and obfuscated logic attacks unless its safety layer is significantly reinforced.

With ChatGPT serving 700 million weekly users, OpenAI maintains its position as the largest player in generative AI by audience. However, this immense scale has brought considerable growing pains. Altman noted that API traffic doubled within 24 hours of the GPT-5 launch, contributing to platform instability. While OpenAI aims to mitigate this by doubling rate limits for ChatGPT Plus users and continually tweaking its infrastructure based on feedback, these early missteps—compounded by confusing user experience changes and high-profile launch errors—have created an opening for competitors to gain ground. The pressure is mounting for OpenAI to demonstrate that GPT-5 represents a substantial leap forward, not just an incremental update. Based on the initial rollout, many users remain unconvinced.