OpenAI: churn and retirements

OpenAI keeps changing the parts around ChatGPT faster than most companies change a homepage. In one week of April 2026, the clearest public map of those changes was still a help-center page, a batch of older Codex models got a retirement date of April 14, and the company’s model-behavior work was still being reshaped after a 2025 reorganization. That is not one headline. It is a pattern. (help.openai.com) The first piece of that pattern is where OpenAI tells people what changed. OpenAI’s ChatGPT release notes page was updated on April 2, 2026, and it continues to function as the main public changelog for product changes like Apple CarPlay support, updated app connectors, mobile interface changes, location sharing, and new Codex plugin features. The company does publish blog posts and research pages, but the help-center notes are still where many day-to-day product shifts first appear in one running timeline. (help.openai.com) That matters because OpenAI is no longer changing one chatbot with one model. The April 2026 release notes describe ChatGPT as a bundle of moving parts: apps, voice, mobile layouts, location controls, Codex workflows, and model routing. A user can feel like they are using one product, while the underlying model, interface, and connected tools change underneath them on different schedules. (help.openai.com) The second piece is model retirement. On April 8, 2026, an OpenAI Codex GitHub discussion announced that older models would be retired in Codex for people signing in with a ChatGPT account starting April 14. The list includes gpt-5.2-codex, gpt-5.1-codex-mini, gpt-5.1-codex-max, gpt-5.1-codex, gpt-5.1, and gpt-5. (github.com) OpenAI is not removing everything at once. It is narrowing the menu. After April 14, the same Codex post says ChatGPT-sign-in users will be pointed to gpt-5.4, gpt-5.4-mini, gpt-5.3-codex, gpt-5.3-codex-spark for Pro users, and gpt-5.2. The message is simple: fewer legacy options, more pressure toward the newer default stack. (github.com) This is not an isolated cleanup. OpenAI had already retired GPT-5.1 models in ChatGPT on March 11, 2026, according to its model release notes. Before that, OpenAI announced on January 29, 2026 that GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini would be retired from ChatGPT on February 13, 2026, while API access would remain unchanged. (help.openai.com) That split between ChatGPT and the application programming interface is one of the easiest places for teams to get tripped up. OpenAI’s January 29 post explicitly said the February 13 retirements applied to ChatGPT and that “in the API, there are no changes at this time.” Its Codex help page also says model availability in ChatGPT is separate from Codex and that application programming interface availability and pricing are managed separately. A model can disappear from one surface and still exist in another. (openai.com) The third piece of the story is the people deciding how these systems behave. In September 2025, TechCrunch reported that OpenAI reorganized its Model Behavior team, a group of about 14 researchers focused on how models interact with people. The team was moved under Post Training lead Max Schwarzer, and founding leader Joanne Jang shifted to a new internal project called OAI Labs. (techcrunch.com) That team’s work sits unusually close to product decisions users actually notice. TechCrunch reported that the Model Behavior team worked on personality, political bias, and reducing sycophancy, which is the tendency for a system to flatter or agree too easily. OpenAI’s own January 29 post on retiring older ChatGPT models said user feedback on GPT-4o’s warmth helped shape GPT-5.1 and GPT-5.2, and it promised more work on personality, creativity, unnecessary refusals, and overly cautious replies. (techcrunch.com) OpenAI is also making that behavior work more public than before, at least on paper. On March 25, 2026, OpenAI published “Inside our approach to the Model Spec,” describing the Model Spec as a public framework for model behavior that balances safety, user freedom, and accountability. That publication arrived days after other safety-related research posts, including one on monitoring internal coding agents for misalignment on March 19. (openai.com) Put those pieces together and the operational picture looks less like a stable platform and more like a fast-moving transit system. The release notes keep changing. The model list keeps shrinking and rerouting. The teams responsible for behavior and safety keep getting reorganized and folded closer to core development. Even when OpenAI explains the changes clearly, the changes are still constant. (help.openai.com) For companies building on OpenAI, the practical lesson is not to treat any single model name like a permanent foundation. If your product assumes one exact model, one exact tone, or one exact tool path, every retirement notice becomes an outage risk. The safer design is modular: keep prompts versioned, isolate model-specific behavior behind a switch, test replacements before deadlines, and assume the chat product, coding product, and application programming interface may diverge. That conclusion is an inference from OpenAI’s recent retirement notices, release-note cadence, and team changes. (help.openai.com) OpenAI is still shipping quickly. But the bigger story in April 2026 is not one launch. It is churn: models aging out in weeks, policies being rewritten in public, and the people shaping model behavior moving inside a structure that is still being rearranged. If you build on top of that, replaceability stops being a nice architecture principle and starts looking like basic hygiene. (github.com)