gpt-image-1.5 deprecation: December 1 is the cutoff, and gpt-image-2 is not the replacement

OpenAI's deprecations page lists the gpt-image-1.5 deprecation for December 1, 2026, alongside gpt-image-1-mini and chatgpt-image-latest. The recommended replacement is not gpt-image-2, which is what most coverage says. The table names gpt-image-2.5-sunburst or gpt-image-2.5-flare. If your code still sends any of those three model ids, it needs a change this month, and the parameters change with it.

What the gpt-image-1.5 deprecation table says, row by row

The notice is dated June 2, 2026, and OpenAI's wording is "deprecation and removal from the API on December 1, 2026". This is the table as of October 31, 2026, with the already-retired gpt-image-1 row added from the same page.

Two details in that table are easy to miss. The dated snapshot gpt-image-1.5-2025-12-16 is listed on the same row as the alias, so pinning the snapshot buys you no extra time. And gpt-image-1 is not an upcoming deprecation at all: the page puts it under past deprecations with an October 23, 2026 shutdown, and the image generation guide repeats that it "was shut down on October 23, 2026". If a service of yours still calls gpt-image-1, it is already failing.

The notice period matches OpenAI's stated policy of at least six months for generally available models, which the same page defines. June 2 to December 1 is one day short of six months.

Which OpenAI image models are still supported

The Create image reference says the supported models for the model parameter include nine GPT Image ids: gpt-image-1-mini, gpt-image-1.5, gpt-image-2, gpt-image-2-2026-04-21, gpt-image-2.5-sunburst, gpt-image-2.5-sunburst-2026-09-08, gpt-image-2.5-flare, gpt-image-2.5-flare-2026-09-08 and chatgpt-image-latest. Three of those nine disappear on December 1, leaving the gpt-image-2 and GPT Image 2.5 families.

Between the two replacements, the docs are specific: choose Sunburst "for workflows where editing precision matters most" and Flare for "fast, high-quality everyday image generation". Their default snapshots are gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08.

How to find the model id in your code before December 1

1. Search your repositories, environment files and prompt templates for five strings: gpt-image-1.5, gpt-image-1.5-2025-12-16, gpt-image-1-mini, chatgpt-image-latest and gpt-image-1.
2. Check both places the id is set. With the Image API you pass it as the model on /v1/images/generations and /v1/images/edits. With the Responses API you pick a mainline model at the top level and name the image model in the image generation tool's own model field, so a grep of your top-level model config alone will miss it.
3. Ask the API instead of guessing. The deprecations page documents an optional shutdown_date field on GET /v1/models and on the retrieve-model endpoint, returned as a YYYY-MM-DD string or null when no shutdown has been announced. That is one assertion you can add to a test and leave running.
4. Check queued Batch work. Both the old and new image models support /v1/batch, so a job that sits in a queue across the cutoff is worth draining early.

What else changes with the model id

The swap is not purely a string replacement. From the image generation guide and the image edit reference, as of October 2026:

• Quality gains two levels. Sunburst and Flare accept low, medium, high, xhigh, max and auto. Earlier GPT Image models stop at high. Both default to auto, so an unspecified quality may not cost what it used to.
• Sizes are freer, with rules. Both 2.5 models take arbitrary WIDTHxHEIGHT strings such as 1536x864, provided both edges are multiples of 16, the aspect ratio is between 1:3 and 3:1, the total pixel count is between 655,360 and 8,294,400, and no edge exceeds 3840 pixels. Anything above 2560x1440 is marked experimental.
• input_fidelity is the parameter to test. The edit reference says it supports high and low on gpt-image-1.5, only low on gpt-image-1-mini, and should be omitted for gpt-image-2. It does not state behavior for the 2.5 models, so verify what your calls do with it rather than assuming the gpt-image-1.5 setting carries over.
• The response can stop containing text. The gpt-image-1.5 and chatgpt-image-latest model pages list output modalities of image and text and price text output at $10 per million tokens. The Sunburst model page lists image output only and says text output is not billed because the model outputs images. If anything downstream reads text from an image response, check it.
• Cached input works differently. For GPT Image 2 and 2.5, cached input pricing applies only to the image generation tool in the Responses API, not to direct Images API calls including /v1/images/edits, and the cached token counts are not returned in the Responses output, so usage cannot confirm a cache hit.
• response_format is gone for these models. The reference calls it a legacy parameter for retired image models and unsupported for GPT Image models, which always return base64.
• user is deprecated in favor of safety_identifier, capped at 128 characters, with a non-empty safety_identifier taking precedence over user.
• One new capability is worth the migration on its own: a markup parameter that overlays a transparent PNG on the first input image to direct an edit. It is documented for GPT Image 2.5 and cannot be combined with mask.

Is gpt-image-2.5 more expensive than gpt-image-1.5?

On standard per-token rates from OpenAI's pricing page, read on October 31, 2026, the answer depends entirely on which model you are leaving. All figures are per million tokens.

Leaving gpt-image-1.5 or chatgpt-image-latest, the image output rate drops from $32.00 to $30.00 and everything else holds. Leaving gpt-image-1-mini is the painful one: image output goes from $8.00 to $30.00 and image input from $2.50 to $8.00, and the table offers no cheap replacement tier. Batch processing halves the image rates to $15.00 output and $4.00 input for the 2.5 models.

The per-token comparison is not a per-image comparison, and we cannot honestly give you one. OpenAI publishes per-image prices for the outgoing models, including $0.009 for a low-quality 1024x1024 image on gpt-image-1.5 and $0.133 for the same size at high quality. The Sunburst and Flare model pages publish no equivalent per-image table, and the guide states plainly that equal token rates do not mean equal cost per image because token consumption differs by model and quality setting, and that the gpt-image-2 calculator does not estimate GPT Image 2.5 consumption. So measure the usage field on your own prompts, sizes and quality settings before December 1 instead of trusting a ratio.

Does this affect ChatGPT or only the API?

Only the API, on the evidence in the docs. The deprecation entry is scoped to "removal from the API", and chatgpt-image-latest is an API alias rather than the ChatGPT app's live model: its model page describes it as pointing to the image snapshot "previously used in ChatGPT" and recommends Sunburst for API use. Nothing in that entry or those model pages says anything about the ChatGPT consumer apps changing on December 1. If you use ChatGPT rather than the API, these rows are not yours to act on.

Where the coverage of this deprecation goes wrong

The disagreement is worth knowing because it will cost someone a second migration. AI Weekly's alert on the three image APIs names the same three model ids and the same December 1 date but says the replacement is gpt-image-2. So do several other trackers. The live table does not; it names the 2.5 pair.

OpenAI's own page contributes to the confusion. The older November 14, 2025 DALL-E entry still recommends migrating to "gpt-image-2, gpt-image-1, or gpt-image-1-mini", and two of those three have since been retired or scheduled for retirement. On a page where entries are stacked by announcement date and never rewritten, the rule is to read the newest dated entry for your model, not the first row that matches your search.

Do this before the end of November

Grep for the five ids, change them to gpt-image-2.5-flare for straight generation or gpt-image-2.5-sunburst for edit-heavy work, drop response_format, move user to safety_identifier, re-test input_fidelity and quality, and log the usage field on a representative sample so you know your real per-image cost rather than a ratio from a rate card.

If the forced change is a reason to look wider than one vendor, Upfynd's AI Image & Video category lists 403 products as of October 2026, among them Stable Diffusion, and the LLMs & AI Infrastructure category lists 275, including Replicate for hosted model inference and PricePerToken for side-by-side API price comparison. December 1 is a crowded date for API bills: it is also when the Gemini API price increase we covered takes effect.