OpenAI Retires o3 and Google Shuts Down Imagen 4: August 2026's Model Retirement Calendar

Every few months, some engineering team discovers in production that the model they'd been using for a year stopped existing overnight. Here are the two dates worth putting on the calendar right now.

OpenAI Retires o3 on August 26

OpenAI retires its o3 reasoning model from ChatGPT on August 26, 2026, after a 90-day sunset period. Both o3 and GPT-4.5 (already retired earlier) were available only to paying subscribers, giving teams that depend on o3's specific reasoning behavior just weeks to evaluate GPT-5 series alternatives and adjust any workflow built around its particular behavior. Separately, OpenAI notified developers on June 11, 2026 who use old GPT-5 and o3 snapshots that those models will be removed from the API on December 11, 2026.

Google Shuts Down Every Imagen 4 Endpoint on August 17

Google's Imagen 4 models were flagged as deprecated starting June 15, 2026, with a shutdown date set for August 17, 2026. From that date on, any application still calling `generate_images()` on Imagen 4 fails with a hard error — no gradual degradation, no additional grace period. Google's recommended migration target for most use cases is `gemini-3.1-flash-image`, within the Gemini image model family.

A Pattern That Repeats Every Few Months

This pace of retirements — every few months, with 60-to-90-day windows — is already the industry norm, not the exception. The three major providers (OpenAI, Google, Anthropic) now publish deprecation calendars in advance, but actively monitoring them remains the responsibility of the team integrating the model.

What It Means for Production Enterprise Architectures

Any architecture that depends on a hardcoded `model_id` scattered across multiple places in the code is, by design, fragile against this retirement pace. The perennial practical recommendation: centralize model configuration in a single place (an environment variable or config service), subscribe to the deprecation changelogs of every provider you use, and treat every announced retirement as a sprint task with a real deadline, not a notification to ignore until something breaks in production.