On October 23, 2026, OpenAI will shut down one of the most symbolic pieces of its recent history: gpt-4o-2024-05-13, the original GPT-4o snapshot the company introduced in May 2024 as its first “omni” model, able to work with text, audio and images. From that date, applications that still call that identifier in the OpenAI API will receive errors. The company’s official model deprecations page recommends migrating to gpt-5.6-sol.
The retirement is no surprise to attentive developers. According to an analysis published by Ctrl Alt Nod, which reviewed OpenAI’s deprecations page on October 11, the shutdown date for that snapshot was announced on April 22, 2026, about six months in advance. The same page states that OpenAI gives at least six months’ notice for generally available models, at least three months for specialized variants and as little as two weeks for preview versions.
The change has financial implications for anyone still relying on that version. According to AI Change Watch, gpt-4o-2024-05-13 costs $5 per million input tokens and $15 per million output tokens, while gpt-5.6-sol costs $4 and $20, respectively. In other words, migrating makes input 20% cheaper but output 33% more expensive, so the net effect on the bill will depend on each application’s mix of token types.
It is not only that snapshot that goes dark on October 23. The same deprecations page lists gpt-4.1-nano, o4-mini and the fine-tuned version ft-o4-mini-2025-04-16 for that day, each with a named substitute. By contrast, the base gpt-4o model and gpt-4o-mini do not appear on the deprecation list as of this writing. According to the model page, the default gpt-4o snapshot is gpt-4o-2024-08-06, and gpt-4o-2024-11-20 is also listed; neither has an announced shutdown date.
That distinction matters because there is a lot of confusion around GPT-4o. In ChatGPT, the model disappeared months ago. According to OpenAI’s Help Center, on February 13, 2026, the app retired GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini, and customers on Business, Enterprise and Edu plans kept GPT-4o in their custom GPTs until April 3, when it was retired across all plans. Conversations switched by default to the GPT-5.3 Instant and GPT-5.4 Thinking and Pro equivalents.
OpenAI clarified at the time that ChatGPT’s voice mode and image generation were not changing with that decision, because, although they are built on a similar model, they are separate systems from the text GPT-4o. In addition, the chatgpt-4o-latest alias, which pointed in the API to the version used in ChatGPT, was shut down on February 17, 2026. The audio and realtime models in the GPT-4o family have their own dates, scheduled for January 2027, and the transcription models for February 2027.
The retirement of GPT-4o in ChatGPT triggered an emotional reaction that is unusual in the software world. Many users felt the model had a personality of its own, warmer or more creative than its successors. According to reports from January cited by several outlets, OpenAI justified the decision on the grounds that only a very small fraction of users were still choosing GPT-4o each day, although that figure could not be verified directly on the announcement page.
With the October 23 date approaching, that movement has resurfaced on X under the #keep4o hashtag. According to the summary of the conversation on that platform, participants are sharing stories about how particular versions of GPT-4o supported them through difficult times or creative projects, pointing to specific snapshots they remember as especially good and asking OpenAI to release the models as open source or keep some form of archival access to preserve them.
The request runs up against business realities. Keeping old models running carries an infrastructure cost, and OpenAI argues that its new models are better and that usage of the old ones is low. Releasing the weights of a commercial model like GPT-4o, meanwhile, would mean handing proprietary technology to competitors, something the company has not proposed. Even so, the campaign reflects a growing tension between the industry’s logic of constant renewal and the bond many people form with a specific model.
For developers, the practical message is more prosaic. Anyone with code that calls gpt-4o-2024-05-13 needs to find and migrate it before October 23. Specialists recommend searching code and configuration files for every reference to dated versions, testing the substitute with representative prompts and measuring quality, latency, cost and output format before switching production traffic, as well as keeping a rollback plan for the most critical workflows.
The case also illustrates a broader dependency problem. Companies that pin a specific snapshot do so precisely to get stable, predictable behavior. When that version is retired, the replacement rarely responds in exactly the same way, and the differences can affect output formats, tone or classification results. OpenAI warns that migrated conversations and GPTs move to the “closest equivalent,” which ensures continuity but not identical behavior.
Beyond nostalgia, the retirement marks the end of an era. GPT-4o was the model with which OpenAI popularized fluid voice conversations and real-time multimodality, and for months it was the default experience for hundreds of millions of ChatGPT users. Two and a half years later, the pace of the industry has turned it into a legacy version, replaced by generations that arrive every few months.
On October 23, the original snapshot will stop responding in the API. The #keep4o campaign is unlikely to change that date, but it has opened a debate the industry will have to face more often: what happens to retired models, who decides whether they are preserved and how the transition is managed for those who depend on them, whether for technical or personal reasons. For now, OpenAI recommends migrating to GPT-5.6 Sol and reviewing its deprecation schedule well in advance.
