GPT Image 1 Shutdown: Product Image Migration Guide

OpenAI shut down `gpt-image-1` on 23 October 2026. If a product-image automation still names that model, this is no longer a future migration task. OpenAI’s current API reference directs users to GPT-Image-2.5 Sunburst or Flare.

Changing the model ID is necessary, but it is not a complete product-photo migration. A new model can interpret the same prompt differently, preserve a reference more or less accurately, use different quality options, and produce a file that behaves differently in your listing workflow.

The safest approach is to keep the old prompt and output requirements fixed, run a small regression set on the replacement model, then retune only the cases that fail. Do not redesign the prompt and change the model at the same time, because you will not know which change caused the result.

Choose the replacement by the job

OpenAI recommends GPT-Image-2.5 Sunburst as the current replacement on the retired model page. Its image-generation guide also positions Sunburst for workflows where editing precision matters most, while Flare is intended for fast, high-quality everyday generation.

For product sellers, that suggests two routes. Use Flare for rapid concepts, social variations and low-consequence drafts where speed matters. Use Sunburst for approved-product edits, polished campaign images, label-sensitive work and any output where a quiet product change creates real risk.

This does not mean every final image must use the slower model. A better rule is to route by consequence. A background concept can tolerate a rejected draft. A medical ID engraving, stone setting or branded package cannot.

The Lustra article on the Flare and Sunburst two-pass workflow explains how to shortlist with Flare and move only approved directions to Sunburst. A migration audit should first prove that each route is safe for its assigned job.

Build a five-case regression set

Do not test with one attractive product photo. Choose cases that represent the ways your workflow can fail:

A clean catalogue master where silhouette, colour and hardware must remain unchanged.
A background replacement that should alter the scene but not the product.
A text-sensitive item such as packaging, an engraving or a printed label.
A transparent-background asset with fine edges, holes, glass or chain links.
A multi-turn edit where the second or third instruction must not degrade the first approved result.

Use the same source images and prompts for both the retired workflow records and the replacement test. If you did not save old outputs, compare against the real product and the approved listing asset instead of relying on memory.

Freeze output settings before judging quality

OpenAI’s current image-generation guide lets applications set model, quality, size, format, compression and background. GPT Image 2.5 adds `xhigh` and `max` quality options, supports custom dimensions within documented limits and can return transparent PNG or WebP files.

Set these values explicitly during migration. Leaving size, quality or background on `auto` may be convenient, but it makes before-and-after results harder to compare. A sharper output may simply have more pixels. A slower result may be using a higher quality level. A missing alpha channel may be an output-setting problem rather than a model failure.

Record at least the chosen model, quality, dimensions, file format, background mode and whether the request was a generation or an edit. Measure processing time and rejected-output rate on your own workload. Do not assume equal token rates mean equal cost per approved image, because retries and manual review can dominate.

Test product fidelity separately from visual quality

A replacement image can look more realistic and still be less usable. Score each case in two columns.

The first column is visual quality: lighting, texture, composition, edge quality and overall polish. The second is product fidelity: shape, proportions, colour, material, logo, text, hardware, stone count and anything the buyer will receive.

A beautiful image with a changed clasp fails the product column. A perfectly preserved product with a poor shadow may still be repairable. Keeping the scores separate prevents aesthetics from hiding a catalogue error.

OpenAI says Images 2.5 improves reference-subject preservation and precise editing, but it does not promise exact reproduction of every small product detail. The guide to reference fidelity in ChatGPT Images 2.5 provides a jewellery-specific checklist for clasps, links, engravings and stone settings.

Use the right API shape

OpenAI distinguishes between the Image API and the Responses API. The Image API suits a single generation or edit from one request. The Responses API is designed for conversational, multi-turn image workflows and can keep image inputs and outputs in context.

Do not migrate endpoints merely because you are migrating models. If a single-edit workflow already matches your product process, keep that structure while validating the replacement. Move to a multi-turn flow only when it solves a real need, such as staged approval or iterative local changes.

Every extra turn is another chance for drift. Save the approved source and compare each iteration with it, not only with the previous AI output.

Retune prompts after the baseline

OpenAI’s prompting guide recommends describing the subject, composition, style and constraints, and specifying what must remain unchanged during edits. Start the migration with the current prompt so you can measure the model change. Then simplify or strengthen only the instructions linked to an observed failure.

For example, if a bracelet keeps its shape but changes the clasp, add a precise preservation instruction for the clasp and supply a clearer reference view. If text fails, keep final legal or packaging copy as editable artwork rather than repeatedly asking the image model to render it.

Avoid adding a long list of generic negative instructions to every request. More words do not guarantee more control. Product facts, allowed changes and locked details should be easy to distinguish.

A release checklist for product-image pipelines

Replace every explicit `gpt-image-1` model reference and check scheduled or rarely used jobs too.
Run the five-case regression set with fixed output parameters.
Choose Flare or Sunburst by the consequence of a failed image.
Compare visual quality and product fidelity as separate scores.
Verify file dimensions, format, compression and transparency in the destination channel.
Measure approval rate, latency and retries before changing production defaults.
Keep the real product image as the permanent source of truth.

The practical conclusion

GPT Image 1 has already stopped, so waiting is not a valid migration plan. The useful work is not a blind model-name swap. It is a controlled test that proves the replacement preserves the product, produces the expected file and fits the same approval process.

Migrate the simplest cases first, keep output settings explicit and route high-consequence edits to the more precise model. Once the baseline passes, prompt improvements become optimisation rather than emergency repair.

Sources

OpenAI API: GPT-Image-1 model page, accessed 2 November 2026.

OpenAI API: Image generation guide, accessed 2 November 2026.

OpenAI API: Image prompting guide, accessed 2 November 2026.

OpenAI: Introducing ChatGPT Images 2.5, 8 September 2026.