Google shut down Imagen 4’s Gemini API endpoints on August 17, 2026, retiring the
imagen-4.0-generate-001, imagen-4.0-fast-generate-001, and imagen-4.0-ultra-generate-001
models, per Google’s own Gemini API release notes and changelog.
Developers still calling those endpoints after the cutoff get hard failures, not a graceful fallback.
What actually changed
Google’s documented replacement is Gemini 3.1 Flash Image, the model most developers know by its
codename, Nano Banana 2, but the migration is not a drop-in model swap. The generate_images()
method that Imagen calls used is gone entirely for the replacement model; image generation now runs
through generate_content(), the same call a developer would already use for text generation,
according to reporting from Virtual Arc Software
and Clauding, which both walked through
the changed request shape after the endpoints went dark.
This is the second stage of a staged retirement. Google deprecated Imagen 4 on Vertex AI back on March 24, 2026, with that platform’s own endpoints hitting their deprecation date on June 30. The Gemini API’s Imagen 4 endpoints, the ones most individual developers and smaller integrations actually called, were the last piece still standing until this week’s cutoff.
Why it matters
A method-signature change, not just a model-ID swap, is the kind of migration that breaks silently in
production if a team only updates the model string and assumes the rest of the call still works. Any
integration built directly against generate_images() needs a real code change, not a config edit,
before it can call the replacement model at all.
The four-month gap between the Vertex AI deprecation and this week’s Gemini API shutdown likely gave most production teams enough lead time to migrate deliberately, but it also means any integration that was quietly ignoring deprecation warnings has now had its assumption broken all at once, on a specific date, with no soft landing.
What to watch: whether Google applies the same generate_content()-first pattern to future image and
video model releases, which would make this migration a one-time adjustment rather than a recurring
one every time an image model gets replaced.
