OpenAI listed gpt-image-1 as a legacy snapshot on 22 April 2026, with a shutdown date of 23 October 2026 and two named replacements, gpt-image-2.5-sunburst and gpt-image-2.5-flare. The deprecations page says generally available models get at least six months of notice, and this one got exactly that.
On 24 October my image generation tool came back with an HTTP 400 saying the model did not exist.
Four callers, one variable
My ops server has four tools that call the image API. Here is how each one chose its model, and where each stood after the shutdown.
| Tool | How it picks a model | After 23 October |
|---|---|---|
| Hero image | Reads a variable, defaults to gpt-image-2 | Unaffected by the shutdown |
| Generic image | Named gpt-image-1 in the request body | Broke, now reads a variable |
| Ad photo | Named gpt-image-1 in the request body | Still broken on 24 October |
| Image edit | Named gpt-image-1 in the request form | Still broken on 24 October |
Three of four had the retired name written into the code. Two of four are still down as I write this, because I fixed the one I needed. That is a fair picture of how these things go.
The notice was never the problem
Every guide tells you to watch the deprecations page, and I could have. Six months is plenty. What I did not have was a list of everything that names a model, and the places most likely to hold an old name are the ones nobody thinks of as products: internal scripts, ops helpers, the glue you wrote on a Tuesday.
A deprecation notice only helps if you can answer “where do we name this model?” in under a minute.
The check I run now
Search every repository and tool directory for model strings, including the folders you would not call a product:
grep -rnE "gpt-|claude-|gemini-" --include="*.py" --include="*.ts" --include="*.json" .
Write the hits into a table with an owner and a shutdown date if one exists. Then give each tool family one variable with a default and an environment override. The fix on my generic tool was a single line. Finding out which tools needed it was the actual work.
I did not add a fallback to an older model, because once a name is retired there is no older model to fall back to. The only fallback that exists is a setting you can change in one place, which is why the variable matters more than any retry logic.
Retiring a model is a routine event, and it should read as a calendar entry. It becomes an incident only when you cannot see where the name lives.