OpenAI retires older models
OpenAI is removing legacy models from its developer tooling, starting with model retirements in Codex on April 14 — a move that signals continuing consolidation of its model lineup rather than a single big “GPT‑5” launch. (thetechoutlook.com). Full‑stack retirements already reported in February include several named models — GPT‑5 (Instant and Thinking), GPT‑4o, GPT‑4.1, GPT‑4.1 mini, and o4‑mini — which OpenAI pulled from ChatGPT earlier this year. (thetechoutlook.com).
# OpenAI retires older models OpenAI is continuing a cleanup of its model lineup, and the latest step starts in Codex on April 14, 2026. The change is not a single dramatic “GPT-5 launch” moment. It looks more like a steady retirement plan: older models are being removed from ChatGPT and developer tools while newer families take their place. (developers.openai.com, openai.com) This is easiest to understand if you think of OpenAI’s products as two connected layers. One layer is ChatGPT, where people pick models in a chat interface. The other is the application programming interface, or API, where developers build apps and workflows on top of OpenAI models. OpenAI has been retiring models in one layer without always removing them from the other at the same time. (openai.com, help.openai.com) That distinction matters because some of the reporting around these retirements blends the two together. OpenAI’s own January 29, 2026 announcement said that GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini would be retired from ChatGPT on February 13, 2026, alongside a previously announced retirement of GPT-5 Instant, GPT-5 Thinking, and GPT-5 Pro. In the same announcement, OpenAI explicitly said “In the API, there are no changes at this time.” (openai.com) OpenAI’s Help Center repeats the same point even more directly. Its support article says GPT-4o and additional models were deprecated in ChatGPT on February 13, 2026 and that these models continue to be available in the API. That means the February change was a product-interface retirement, not a full shutdown across every OpenAI surface. (help.openai.com) There was one temporary exception for business customers. OpenAI says ChatGPT Business, Enterprise, and Edu customers kept access to GPT-4o inside Custom GPTs until April 3, 2026, after which GPT-4o was fully retired across ChatGPT plans. Even there, OpenAI still says API access remains unchanged. (help.openai.com, help.openai.com) The newer April move affects Codex, which is OpenAI’s coding product for the command line, integrated development environments, and its standalone app. OpenAI’s developer documentation now has a deprecations page that lists model shutdowns and replacement paths, including Codex-related removals. One entry says developers using the codex-mini-latest snapshot were notified on November 17, 2025 that it would be removed from the API on February 12, 2026, and that OpenAI would also stop supporting its legacy local shell tool tied to that model. (developers.openai.com) Codex itself is also being updated around newer model families rather than preserved as a museum of old ones. OpenAI’s Codex changelog says the company introduced gpt-5-codex-mini in March 2026 as a smaller, cheaper option, and OpenAI’s model release notes describe it as a lighter version of GPT-5-Codex that gives users up to 4 times more usage in ChatGPT. That is the pattern behind these retirements: remove older choices, steer users toward a smaller set of current defaults, and offer cheaper “mini” versions where needed. (developers.openai.com, help.openai.com) OpenAI’s own wording makes the strategy unusually clear. In the January retirement post, the company said retiring models is difficult but lets it focus on improving the models most people use today. That is a consolidation strategy, not a sign that one giant flagship model will simply replace every previous product overnight. (openai.com) This also helps explain why the “GPT-5” label has been confusing. OpenAI’s public materials in early 2026 refer to multiple GPT-5 variants, including Instant, Thinking, Pro, and Codex-specific versions such as GPT-5-Codex and GPT-5-Codex-Mini. Instead of one clean handoff from “GPT-4” to “GPT-5,” OpenAI appears to be running a family of specialized models and trimming older branches as new ones become good enough and cheap enough to support at scale. (openai.com, help.openai.com, developers.openai.com) There is another clue in OpenAI’s current Codex help documentation. The company says users can still use a model picker and check “Legacy” for older versions in Codex, but it also warns that model availability in ChatGPT is separate from Codex and may change over time, and that API availability and pricing are managed separately. In other words, OpenAI is actively managing access by product surface, not treating every model as universally available everywhere. (help.openai.com) For developers, the practical takeaway is simple. A model disappearing from ChatGPT does not automatically mean it is gone from the API, and a Codex retirement does not necessarily mean every OpenAI coding workflow is being shut down. The important thing is to check the deprecations page, the relevant help article, and the replacement model OpenAI recommends for that specific tool. (developers.openai.com, help.openai.com) For everyone else, this is a sign that OpenAI is becoming more like a conventional software platform. Mature platforms do not keep every old version around forever. They narrow the lineup, move users to current defaults, and keep the newest tools where they think performance, reliability, and cost are easiest to manage. OpenAI’s 2026 model retirements fit that pattern closely. (openai.com, developers.openai.com)
