OpenAI Image Generation API Models: Which One Should You Use in 2026?

As of October 9, 2026, the OpenAI API lists exactly three live image generation models on its All models page: gpt-image-2.5-sunburst, gpt-image-2.5-flare, and gpt-image-2. The latest ones are the two GPT-Image-2.5 models, both with a 2026-09-08 snapshot. Pick gpt-image-2.5-sunburst when editing precision matters, gpt-image-2.5-flare when you want fast everyday generation, and use gpt-image-2 if you already run it or want the lowest published per-image price ($0.005–$0.006 at low quality); it is not deprecated.

If the ID in your code is gpt-image-1, it stops working on October 23, 2026. If it is gpt-image-1.5, gpt-image-1-mini, or chatgpt-image-latest, the shutdown date is December 1, 2026. OpenAI's deprecations page names gpt-image-2.5-sunburst or gpt-image-2.5-flare as the replacement for all four. dall-e-2 and dall-e-3 were shut down on May 12, 2026.

Below, the full ID table with status, dates, and token rates comes first, then a decision rule by job, the exact official page for each fact, and the distinction that causes the most confusion: image recognition in the OpenAI API is done by mainline models such as gpt-5.5, not by GPT Image models.

GPT Image models in the OpenAI API: every ID, status, and shutdown date

The table pulls together what OpenAI spreads across the All models page, the individual model pages, the deprecations page, and the pricing page, as of October 9, 2026. Token rates are per 1M tokens on the Standard tier; image output is the rate that dominates a per-image bill.

Three details change how you read that table:

• The API reference still lists retired IDs. The model enum on the Create image reference includes dall-e-2 and dall-e-3 even though both stopped serving on May 12, 2026. Treat the All models page and the deprecations page as the source of truth for what is actually alive, not the enum.
• Text tokens are priced the same on the three live models: $5 per 1M input, $1.25 per 1M cached input. The deprecated models differ on text too (gpt-image-1-mini is $2 / $0.20; gpt-image-1.5 and chatgpt-image-latest also charge $10 per 1M text output).
• Batch halves the image rates. On the three live models, Batch image tokens are $4 in / $1 cached / $15 out per 1M. Cached input rates for GPT Image 2 and 2.5 apply only to images generated through the Responses API, not through direct /v1/images/* calls, per the pricing page footnote.

Per-image dollar prices exist for only one of the three live models. The image generation guide's "Calculating costs" section publishes this table for gpt-image-2:

For gpt-image-2.5-sunburst and gpt-image-2.5-flare there is no static table, only the guide's interactive calculator.

All three live models expose the same endpoints (v1/images/generations, v1/images/edits, v1/batch), support inpainting, and ship with the same default rate limits: 250,000 tokens and 20 images per minute on the Build tier, 3,000,000 and 150 on Launch, 8,000,000 and 250 on Grow.

Latest OpenAI image generation model: GPT-Image-2.5 Sunburst and Flare

"Latest" has a precise meaning here. Both 2.5 models have a single dated snapshot, 2026-09-08, and the alias IDs gpt-image-2.5-sunburst and gpt-image-2.5-flare resolve to it by default. The community announcement "Introducing GPT Images 2.5 in the API and ChatGPT" is dated September 8, 2026, which matches the snapshot. gpt-image-2's only snapshot is gpt-image-2-2026-04-21, and OpenAI's "Introducing ChatGPT Images 2.0" post carries the same April 21, 2026 date.

What the 2.5 generation adds over earlier GPT Image models, according to the image generation guide and the model pages:

• Two more quality steps. quality accepts low, medium, high, xhigh, max, and auto on both 2.5 models. Earlier GPT Image models, including gpt-image-2, stop at high.
• A default on edits. /v1/images/edits defaults to gpt-image-2.5-sunburst when you omit model. /v1/images/generations has no default; the reference tells you to specify a model explicitly.
• Two positions instead of one headline model. The guide now names only two current models and splits them by job: Sunburst where "editing precision matters most", Flare as "our fastest model for high-quality, everyday image generation".

If you pin snapshots in production, use the dated IDs (gpt-image-2.5-sunburst-2026-09-08, gpt-image-2.5-flare-2026-09-08). If you want to follow whatever OpenAI promotes as current, use the aliases and watch the deprecations page.

Best OpenAI model for image generation: pick by job, not by name

The honest short answer to "which OpenAI model is best for image generation" is that the three live models share identical token rates, so the choice is about fit and about how many tokens each one spends per image, not about a cheaper list price. Only gpt-image-2 turns that into a published per-image price.

Why equal token rates do not mean equal cost per image

The guide states this directly: "Equal token rates don't mean equal cost per image." Cost comes from how many output image tokens a model spends for a given size and quality. OpenAI publishes those counts for earlier models: 272 tokens at low, 1,056 at medium, and 4,160 at high for a 1024×1024 image. Multiply by the output rate to get the image cost:

• gpt-image-1.5 or chatgpt-image-latest at medium: 1,056 tokens × $32 per 1M = $0.034, which matches the $0.034 per-image price printed on the chatgpt-image-latest model page for a 1024×1024 medium image. At high: 4,160 × $32 per 1M = $0.133, again matching the published $0.133.
• gpt-image-2 spends a different number of tokens at the same rate as the 2.5 models. Working back from its published prices: $0.053 ÷ $30 per 1M ≈ 1,767 output tokens for a 1024×1024 medium image, and $0.006 ÷ $30 per 1M = 200 at low. Same $30 rate, very different bill depending on quality.
• For the 2.5 models, OpenAI has not published token counts or a per-image table, and the Sunburst model page says the GPT Image 2 calculator "does not estimate GPT Image 2.5 token consumption." As an illustration only: if a 2.5 image spent the same 1,056 tokens at medium, it would cost 1,056 × $30 per 1M = about $0.032. The real number may differ, especially at xhigh and max.

The practical way to settle Sunburst versus Flare versus GPT Image 2 for your own prompts: generate a small batch with each, read usage.output_tokens from every response (the Images API returns usage with input_tokens, output_tokens, and a split of image and text input tokens), and multiply by $30 per 1M. Add prompt text tokens at $5 per 1M; for most prompts they are a rounding error next to the image tokens. For a deeper side-by-side with request examples, see GPT Image 2.5: Flare vs. Sunburst, Pricing, and API Examples; for per-image budgeting on the older generation, see GPT Image 2 API Pricing: Cost per Image and Project Budget.

Deprecated GPT Image models: gpt-image-1, 1.5, 1-mini, chatgpt-image-latest

These four IDs still answer requests today, which is why so much existing code still uses them. The dates on the deprecations page:

• gpt-image-1: deprecation announced April 22, 2026; shutdown October 23, 2026.
• gpt-image-1.5, gpt-image-1-mini, chatgpt-image-latest: announced June 2, 2026; shutdown December 1, 2026.
• Recommended replacement for all four: gpt-image-2.5-sunburst or gpt-image-2.5-flare.

What actually changes when you swap the ID:

• Coming from gpt-image-1.5 or gpt-image-1: the request shape is the same, and the 2.5 image output rate is lower ($30 per 1M versus $32 for gpt-image-1.5 and $40 for gpt-image-1). You gain xhigh and max. Whether your bill goes up or down depends on token consumption per image, so run the usage check above before and after.
• Coming from gpt-image-1-mini: budget for a rate increase on 2.5. Mini bills image output at $8 per 1M; the 2.5 models bill $30, which is 3.75 times the rate. If low cost was the reason you chose mini, test gpt-image-2 at low first: $0.005–$0.006 per image on the official table, and it has no shutdown date. Also drop input_fidelity on edits: mini only accepted low, and the guide says to omit the parameter for the GPT Image 2 family, which always uses high-fidelity input. The gpt-image-1-mini API page covers that model's request pattern if you need to read the old code before migrating.
• Coming from chatgpt-image-latest: this alias "points to the Image snapshot previously used in ChatGPT", and its model page already recommends GPT-Image-2.5 Sunburst for API use. Its published per-image prices ($0.009 / $0.034 / $0.133 for a 1024×1024 image at low / medium / high; $0.013 / $0.05 / $0.20 at 1024×1536 or 1536×1024) sit alongside the gpt-image-2 table above as the two published baselines; compare both against your own 2.5 measurements, since the 2.5 models have no per-image table.

Nothing on the deprecations page lists gpt-image-2 as deprecated. If you are on it and the April 2026 snapshot does what you need, you can stay; just plan around the fact that OpenAI's guide documents only the 2.5 models as current.

Official OpenAI image generation API docs: which page answers what

OpenAI's official image generation API documentation is spread across six pages, and none of them contains the whole picture. This is where each fact lives:

Two more pages matter for adjacent questions: the Responses image generation tool guide for multi-turn generation inside a conversation, and the Images and vision guide for the recognition side covered next.

OpenAI model for image recognition: use a mainline model, not GPT Image

If your task is to read an image (classify it, extract text, describe a screenshot, answer questions about a photo), you do not use a GPT Image model at all. GPT Image models produce images. The Sunburst, Flare, and GPT Image 2 model pages list only v1/images/generations, v1/images/edits, and v1/batch as supported endpoints; a GPT Image ID is not accepted as the top-level model of a Responses or Chat Completions request.

Image recognition in the OpenAI API means sending the image as input to a mainline language model through the Responses API or Chat Completions. The Images and vision guide lists these models for image input as of October 9, 2026: gpt-6-astra, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.2, gpt-5.1, gpt-4.1, gpt-4.1-mini, gpt-4o, and gpt-4o-mini. The same page marks gpt-5, gpt-5-mini, gpt-5-nano, gpt-4.1-nano, o1, o1-pro, o3, and o4-mini as deprecated for this use.

So the two-sentence routing rule is: text or image in, text out → a mainline model from that list. Text or image in, image out → gpt-image-2.5-sunburst, gpt-image-2.5-flare, or gpt-image-2. A single Responses request can do both, by giving a mainline model the image_generation tool, which is the next section.

OpenAI image generation API: model choice vs endpoint choice

After you have picked a model ID, there are still two places it can go, and putting it in the wrong field is the most common reason a correct ID "does not exist".

Images API (POST /v1/images/generations and POST /v1/images/edits). The guide's rule: "If you only need to generate or edit a single image from one prompt, the Image API is your best choice." The GPT Image model ID goes straight into model. GPT Image models always return the image inline as base64 in data[].b64_json; url is unsupported for them, and there is no endpoint to fetch an image later by ID, so your code has to decode and store the bytes itself. Edits accept up to 16 input images and an optional mask. The old /v1/images/variations endpoint is retired; use edits with a prompt instead. Ready-to-copy requests are in OpenAI Image Generation API cURL: Current Generate and Edit Requests.

Responses API with the image_generation tool. For conversations, multi-step edits, or generating an image as part of a larger agent flow, the top-level model is a mainline text model and the GPT Image model goes inside the tool:

{
  "model": "gpt-5.5",
  "input": "Draw a line-art icon of a paper plane, then make it blue",
  "tools": [
    { "type": "image_generation", "model": "gpt-image-2.5-sunburst" }
  ]
}

The tool-level model accepts gpt-image-2.5-sunburst, gpt-image-2.5-flare, gpt-image-2, and, until their shutdown dates, gpt-image-1.5, gpt-image-1, and gpt-image-1-mini. For the top-level model, the guide says image generation in Responses "requires gpt-5 or a newer model"; the tool page lists a wider set, so pick a current mainline model and confirm support on its model page before you rely on it. The image comes back as an image_generation_call item whose result is base64; pass previous_response_id to keep editing the same image across turns. Two billing consequences: Responses requests also bill the mainline model's tokens, and this is the only path where the $2 cached image input rate applies. Each streamed partial image adds 100 output tokens. The endpoint-level comparison, including what the Responses ig_ IDs are and are not, is in OpenAI Image Generation API Endpoint; streaming previews are covered in OpenAI Image Generation API Streaming: How Partial Image Previews Work.

GPT Image request fails: check shutdown date, field, and parameters first

Most "the model does not exist" or "invalid parameter" failures around these IDs come from one of five things, in this order:

1. The ID has passed its shutdown date. gpt-image-1 after October 23, 2026, or any of gpt-image-1.5, gpt-image-1-mini, chatgpt-image-latest after December 1, 2026. dall-e-3 and dall-e-2 already fail. Switch to a 2.5 model.
2. The ID is in the wrong field. A GPT Image ID at the top level of a Responses request is rejected; it belongs in tools[].model. Conversely, the model enum on /v1/images/generations contains only image model IDs, so a mainline model ID does not work there.
3. A quality value the model does not support. xhigh and max exist only on gpt-image-2.5-sunburst and gpt-image-2.5-flare. Earlier models accept low, medium, high, and auto.
4. A parameter the model rejects. input_fidelity should be omitted for the GPT Image 2 family; gpt-image-1-mini only accepted low. Free WIDTHxHEIGHT sizes must respect the 3,840 px edge limit, multiples of 16, the 3:1 ratio, and the pixel bounds listed above. background: "transparent" requires png or webp output and is marked "in preview" for gpt-image-2.
5. The request was blocked, not misrouted. A content refusal returns error code moderation_blocked with moderation_details that tell you the categories and whether it was the input or the output stage. Changing the model ID will not fix it.

If the error text mentions organization verification rather than the model or a parameter, that is an account-state problem; follow Fix GPT Image 2's 'Your organization must be verified' Error rather than swapping IDs.

FAQ

Which GPT models do image generation in the OpenAI API?

Three live ones as of October 9, 2026: gpt-image-2.5-sunburst, gpt-image-2.5-flare, and gpt-image-2. Four more still respond but are deprecated: gpt-image-1 until October 23, 2026, and gpt-image-1.5, gpt-image-1-mini, and chatgpt-image-latest until December 1, 2026. Mainline models such as gpt-5.5 can trigger image generation in the Responses API, but the image itself is always rendered by one of the GPT Image models above.

What is the latest OpenAI image generation model?

GPT-Image-2.5, in two variants: gpt-image-2.5-sunburst (most capable, editing precision) and gpt-image-2.5-flare (fastest for everyday generation). Both have the 2026-09-08 snapshot.

Is gpt-image-2 deprecated?

No. As of October 9, 2026, it is listed among the three live image models and does not appear on the deprecations page. Its only snapshot is gpt-image-2-2026-04-21, its token rates are identical to the 2.5 models, and it is the only live model with an official per-image table ($0.005–$0.006 at low, up to $0.211 for a 1024×1024 high-quality image).

Is DALL·E 3 still available in the OpenAI API?

No. dall-e-3 and dall-e-2 were shut down on May 12, 2026. The IDs still appear in the API reference's model enum, but requests no longer succeed. Use a GPT Image model; note that GPT Image models return base64 only and do not return revised_prompt on the Images API.

Which OpenAI model is best for image processing or recognition?

A mainline model with image input through the Responses API or Chat Completions, not a GPT Image model. The Images and vision guide currently lists gpt-6-astra, the gpt-5.6 models, gpt-5.5, the gpt-5.4 family, gpt-5.2, gpt-5.1, gpt-4.1, gpt-4.1-mini, gpt-4o, and gpt-4o-mini for that job.

What image model does ChatGPT use, and can I call it from the API?

The API alias chatgpt-image-latest points to the image snapshot previously used in ChatGPT, not the current one, and it shuts down on December 1, 2026. OpenAI's September 8, 2026 announcement introduced GPT Images 2.5 "in the API and ChatGPT" together, so for API work the model page's own recommendation applies: use gpt-image-2.5-sunburst.