GPT Image 2 Retires in February: Plan Your 2.6 Migration

OpenAI's November 10 deprecation notice gives teams using GPT Image 2 a concrete deadline: the API models gpt-image-2 and gpt-image-2-2026-04-21 are scheduled to shut down on February 10, 2027. The recommended replacements are gpt-image-2.6-sunburst and gpt-image-2.6-flare. OpenAI deprecations

The immediate task is to find the jobs that depend on those exact IDs and prove their replacements can deliver acceptable assets. A new model returning an attractive image is not enough if it changes packaging text, drops transparency, increases retries or produces a file the publishing system cannot use.

Treat this as a production migration with a deadline and an acceptance record. The launch features already have their own GPT Image 2.6 API guide. Here, the decision is how to remove the retiring dependency without introducing a different failure into the image pipeline.

The notice distinguishes announcement from shutdown. November 10 is when the retirement was announced; February 10 is the scheduled service deadline. That gives an engineering team a window for comparison and migration, rather than an instruction to replace every production request immediately.

Source: OpenAI's deprecation register. The table concerns the named API models; it should not be used to infer an app-product shutdown or a different provider's contractual deadline.

Search source code and deployment configuration, then inspect the places where requests are persisted. A queued job may contain its own model name. A support dashboard may replay an old request body. A designer's approved template may be copied into many jobs without passing through the current default.

Give each dependency an owner and a last-known execution date. An unused example file is different from a worker that runs only at month-end. Confirm which code paths actually execute before deciding that a search result is harmless or that a clean repository search proves the service is ready.

Follow one representative request from the user's brief to the asset's final destination. Record input preparation, prompt construction, model call, retry behaviour, decoding, storage, image transformation and publication. The replacement must fit the entire path.

For example, a hypothetical ecommerce workflow might generate a product scene, convert the output to a web format, create several crops and attach the result to a draft product page. A successful generation call does not establish that the crops preserve the product or that the final derivative retains transparency. Those are separate acceptance points.

Distinguish newly generated scenes from edits to supplied product photography. An exploratory mood board can tolerate variation that a catalogue asset cannot. Conversely, a strict requirement to preserve every pixel is inappropriate for a job explicitly commissioned to redesign a scene. Write the requirement before judging the output.

Also distinguish requested formats from actual files. A filename extension or response field is not proof that the bytes decode correctly. The file-output acceptance reference explains how to keep file validity and task completion separate.

Capture a small approved baseline set while the current model is still available. Retain inputs, prompts, settings and acceptance decisions, not just favourite images. Include known failures as well as successes so the comparison tests the problem areas the business already understands. Use assets the team is authorised to process, with access and retention appropriate to the material.

OpenAI positions Sunburst for detailed visual work and Flare for faster iteration. Those descriptions give a starting hypothesis; they do not establish which model will produce your accepted output at the lowest total cost. OpenAI image-generation guide

A useful first split is between exploration and delivery. Exploration values fast feedback and variety. Delivery values compliance with the exact brief, compatibility with the destination and a predictable review burden. A team may choose different models for those stages, but it should test the handoff as well as each stage separately.

Suppose an art director approves a rough composition from an exploratory run. Passing that image to another model for refinement may alter details already approved. The workflow needs to say which features are now fixed and which may change. Without that distinction, a second model can create another attractive option while undoing the first decision.

Evaluate candidates on the same brief and source assets. Hide model labels during visual review where practical, then reveal them for the cost and latency decision. Our guide to identity-blind model evaluation provides a structure for reducing brand expectations in that comparison.

The image guide distinguishes capabilities by API surface. For example, markup edits use the Image API's edits endpoint, while the Responses workflow uses masks for regional edits. New capabilities should not be assumed to appear through an unchanged integration. OpenAI image-generation guide

Review your request builder against the endpoint actually used. Check input ordering, masks or annotations, output options and error handling. Documentation examples can demonstrate a pattern without using the model ID you intend to deploy; inspect that value explicitly when adapting an example.

Start with mechanical checks because they are cheaper to diagnose than subjective quality differences. Can the application decode the response? Does it produce the requested number of assets? Are dimensions, orientation and file type suitable? Does the storage layer preserve the required metadata? Does the downstream crop or compression step change the result?

Then test the output in its destination. Place a transparent cutout against the actual page background. Open an exported graphic in the editor the design team uses. Check a mobile crop at its displayed size. These checks answer different questions from zooming into a full-resolution generation preview.

Keep these tests small enough to rerun when an SDK, image processor or publishing integration changes. They are checks of the production contract, not a one-time beauty contest between models.

For edited product imagery, make the allowed change explicit. A request to replace the background does not authorise a different label, cap, accessory, colourway or quantity. Mark the permitted edit region and the protected product details before reviewing results.

Use the source asset as the reference. Store exact expected text separately, identify distinctive shapes and note whether colour is a strict requirement or merely a visual preference. If the task changes lighting, decide which colour differences are acceptable before interpreting a comparison.

A hypothetical bottle image illustrates the distinction. The new background may satisfy the creative brief while the label's volume changes from one value to another. That output fails the product requirement even if reviewers prefer its lighting. A weighted average of beauty and accuracy can conceal this failure; critical product facts should have an explicit veto.

Review at two scales: the final delivery size and a close inspection of protected details. The first reveals composition and practical legibility. The second reveals errors that may be hidden in a thumbnail but appear in a zoomed product view or later reuse.

Record every attempt within the test budget, including failures. If reviewers choose the best of several outputs for one model and only the first output for another, the comparison no longer measures the same process. Separate first-attempt acceptance from acceptance after the allowed retries.

OpenAI warns that matching quality labels across image models can involve different token counts. A setting called high is therefore not a universal billing unit or a guarantee of equivalent output. Record the actual settings and billed usage for each candidate. OpenAI image-generation guide

Separate three amounts: model charges, production labour and downstream infrastructure. The first comes from request usage and applicable rates. The second includes review and correction time. The third may include storage, transformations or delivery. Combining them without labels makes it difficult to understand why a replacement costs more or less.

A useful calculation is:

Cost per accepted asset = (all generation and editing charges + review and repair labour + allocated processing cost) ÷ accepted assets.

Use all attempts in the numerator. If a run produces no accepted assets, report that outcome and the spend; do not manufacture a meaningful unit cost by dividing by an arbitrary number. Keep currency and labour-rate assumptions visible.

For a clearly hypothetical example, suppose a test incurs $12 in model charges and 90 minutes of review priced at $40 an hour. With 24 accepted assets, the combined model-and-review cost is ($12 + $60) ÷ 24 = $3 per accepted asset, before storage and processing. These are illustrative inputs, not OpenAI prices or measured production results.

That example shows why a small difference in generation price may be less important than review burden. It does not prove that either replacement saves labour. Measure the actual workflow and show reviewers the counts behind any average.

Work backwards from February 10 with an internal completion date that leaves room for failed tests and operational changes. The appropriate buffer depends on the number of workflows and the consequences of interruption. Avoid treating the provider's shutdown date as the first day to discover a forgotten dependency.

A practical sequence is inventory, baseline capture, controlled comparison, limited production traffic and complete cutover. Each stage should have an owner and exit evidence. The transition can be gradual, but the remaining use of retiring IDs should be visible.

Pay particular attention to retries. A job created before cutover may fail later and be replayed with its original model field. Decide whether such jobs finish under their original configuration while supported, migrate with an explicit version change, or return for review. Silent mutation makes results harder to reproduce.

Make publishing idempotent at the application level: a retried job should not attach duplicate assets or accidentally replace an approved image twice. Preserve the original task identifier and asset decision across retries. This is an application property to verify, not a capability to infer from the new model.

Before retirement, the old model may remain a temporary comparison or recovery option. After shutdown, it cannot be the operational plan. Establish a supported alternative, a manual production path or a clear delayed-job response for cases the replacement cannot complete.

The final record should explain why each workflow moved to its selected replacement. Include the task set, model IDs, endpoint, settings, test date, acceptance criteria and observed failures. Preserve enough information to rerun the comparison without requiring someone to remember which preview looked best.

Keep unresolved limitations specific. A workflow might accept scene generation but require manual review for packaging text. Another might support a particular export format but defer a newly introduced one. Those are actionable operating boundaries when they are attached to the relevant job type.

Avoid reporting only an overall success percentage. Show the number of tasks, attempts and accepted assets, and split results by the asset categories that matter. A small sample of easy promotional images cannot establish reliability for a large catalogue of tightly specified products.

Have the creative owner confirm visual requirements and the engineering owner confirm delivery behaviour. Neither perspective substitutes for the other. The migration is complete when the system can produce, review and deliver acceptable work through supported paths, including its expected failure cases.

Digital Applied's AI transformation services connect model selection with the practical work of integration, content review and reliable delivery, so a provider deadline becomes a controlled improvement to the production process.