AI image model news, Oct 3, 2026: what API callers must do

The one dated, actionable image-API change on vendor pages this week is xAI's: grok-imagine-image-quality is retired on November 2, 2026. OpenAI's changelog has no image entry for October, Google's release notes list no new image model for September or October, and Black Forest Labs' and Ideogram's documentation carry no dated entries for the period.

Every row below was read on 2026-10-03 from the vendor's own page. Where a page does not date something, we say so instead of filling the gap. The Sume column is from the Image API docs.

What does each vendor's page show?

Read the table as a scope of what is verifiable. A missing entry means the page we fetched has none, not that nothing happened elsewhere.

Which change needs action by a date?

Only the xAI retirement has a deadline. If your code sends grok-imagine-image-quality, plan the move before November 2: xAI says the traffic will be served by grok-imagine-image-2.0 with quality set to low, so choose the model and the quality you actually want instead of relying on the redirect. Before that, compare outputs, because a redirect to low quality may not match what you tuned your prompts against.

If you call Grok through Sume, the model id comes from the catalog, and a field a row does not list returns 400 unsupported_parameter. Check GET /v1/images/models for the current Grok rows.

What about Google's older Nano Banana?

Google's current image generation page labels the original Nano Banana, gemini-2.5-flash-image, as a legacy model and recommends moving to Nano Banana 2 Lite. It lists Nano Banana 2 Lite as the fastest and cheapest of the family, with 1K output only, and Nano Banana 2 and Pro with 512px, 1K, 2K and 4K. If you still pin the legacy id, test the replacement on your own prompts first, because Lite has no Google Search grounding and no 4K.

Does anything change how you pick a model on Sume?

Not from this week's pages. Sume documents ChatGPT Image 2.5 as openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst, takes up to 16 references on them, and publishes each model's accepted parameters on the catalog. The practical habit is the same as ever: read the catalog row, send only fields it lists, and set quality and size explicitly so a vendor default change cannot move your cost.

For a long-running 4K or high-quality request, expect a 202 job envelope after the 30-second blocking budget and read the result from the job endpoints.

How do you keep up without reading every page?

Keep a short list of four URLs: OpenAI's changelog, Google's Gemini API release notes, xAI's release notes, and the BFL documentation home. Add a calendar item for any retirement date you find, and re-read your catalog response on Sume when a vendor announces a new id. We update this kind of roundup when a page shows a new dated entry; nothing here is a prediction.

One more habit helps with retirements: grep your own code and saved request templates for model ids once a quarter. Slugs are easy to forget in a cron job or an automation that nobody has opened in months, and a retired slug turns into a surprise quality change or an error at the worst time. Writing the id, the date you chose it and the reason in a comment next to it takes a few seconds and makes the next migration a review instead of an archaeology project.