Migrating OpenAI image calls from DALL-E 3 to gpt-image: which parameters change?

## What stays the same The transport is unchanged. It is still `POST /v1/images/generations` with a bearer API key, still a JSON body with `model` and `prompt`, and the SDK entry point is still `client.images.generate(...)`. That surface stability is what makes people underestimate the migration: the call compiles and the shape looks familiar, but several parameter *values* are now rejected, and one response field your code reads has vanished. ## Size DALL-E 3 accepted `1024x1024`, `1792x1024` and `1024x1792`. The gpt-image family uses `1024x1024`, `1536x1024` (landscape), `1024x1536` (portrait) and `auto`, which lets the model pick an aspect ratio from the prompt. Any code that stored `1792x1024` in a config file or a database of user presets is now sending a value the model does not accept. `auto` is genuinely useful for user-driven generation, but it means you cannot assume output dimensions — if your UI reserves a fixed box, pin the size explicitly instead. ## Quality DALL-E 3 had two tiers, `standard` and `hd`. gpt-image has `low`, `medium`, `high` and `auto`. This is more than a rename, because quality is now a direct cost lever: the number of billed image output tokens rises sharply from low to high. The migration decision is therefore a product decision — thumbnails, drafts and previews belong at `low`, and only the asset a user commits to needs `high`. A one-to-one mapping of `hd` → `high` across every call path is the most common way a migration quietly triples the image bill. ## Style and response_format: deleted, not renamed `style: "vivid" | "natural"` was specific to DALL-E 3 and has no successor; steer the look through the prompt instead. `response_format` is also gone, because gpt-image always returns base64 in `data[].b64_json`. Any branch in your code that handled `url` output — download the link, follow the expiry, retry the fetch — is dead and should be removed outright rather than left behind a flag. ## Response-side changes DALL-E 3 rewrote your prompt internally and returned the rewritten text as `revised_prompt` on each data entry; teams logged it for debugging and sometimes displayed it. gpt-image does not return that field, so any code reading it must be deleted, and any dashboard built on it loses its source. In exchange the response carries a `usage` object reporting input and output tokens, which is what you now feed into cost telemetry. ## New parameters worth adopting - `background`: `transparent`, `opaque` or `auto`. Transparent output requires an alpha-capable `output_format` (PNG or WebP), and it removes a whole background-removal step from asset pipelines producing icons, stickers or product cut-outs. - `output_format`: `png`, `jpeg` or `webp`, with `output_compression` tuning the lossy formats. DALL-E always gave you PNG; now storage cost is a parameter. - `moderation`: `auto` or `low`, adjusting how strictly the model's own content filtering applies. - `n` greater than 1 is available; DALL-E 3 only supported a single image per call, so batching patterns that worked around that limit can be simplified. ## The variations endpoint `/v1/images/variations` existed only for DALL-E 2 and produced alternates of an uploaded image with no prompt at all. There is no gpt-image variations model, so that capability moves to `/v1/images/edits`: upload the source image and describe the change you want in a prompt. The result is more controllable, but it is not a drop-in — the caller must now supply intent in words, which usually means a product change, not just a code change. ## A migration checklist 1. Grep for hard-coded `1792`, `1024x1792`, `"hd"`, `"standard"`, `"vivid"`, `"natural"`, `response_format` and `revised_prompt`. 2. Delete the URL-download path; keep only base64 decoding. 3. Re-decide quality per call site rather than mapping tiers mechanically. 4. Wire `usage` into cost tracking, since the per-image flat price no longer exists. 5. Re-test moderation behaviour — the filtering and its error surface are not identical to the old endpoints, so prompts that used to pass may not, and vice versa. 6. Replace variations calls with edit calls, and write the prompt that expresses the intent. ## What interviewers are checking The question rewards someone who has actually done the port. The tell of experience is naming the *silent* breakages — quality tiers that changed cost, `revised_prompt` disappearing from logs, and a variations capability that has no direct replacement — rather than reciting that the endpoint path is unchanged.