OpenAI's image models have changed names, versions and prices several times in eighteen months, and any team that calls them through the API now has several shutdown dates to plan around. GPT Image is OpenAI's family of image generation and editing models, introduced in the API as gpt-image-1 on 23 April 2025 and, as of November 2026, sold as GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare. The family replaced the DALL-E models, which OpenAI shut down in the API on 12 May 2026. This page sets out every version with its dated status, the API identifiers, pricing, the main parameters, the provenance measures OpenAI states, and what to check before a shutdown date passes. It is written for developers, product owners, content teams and the security and procurement staff who approve AI tools.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Key takeaways
- Current models are 2.6 Sunburst and 2.6 Flare. OpenAI's deprecations page names gpt-image-2.6-sunburst or gpt-image-2.6-flare as the replacement for every GPT Image model still being retired.
- Three shutdown dates are still ahead or recent. gpt-image-1 shut down on 23 October 2026; gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest shut down on 1 December 2026; gpt-image-2 shuts down on 10 February 2027.
- Pricing is per token, not per image. The 2.6 models cost $5 per 1M text input tokens, $8 per 1M image input tokens and $30 per 1M image output tokens, with Batch at half those rates.
- Access may need verification. OpenAI says organisations may need to complete API Organization Verification before they can use GPT Image models.
- Outputs carry provenance signals. OpenAI states it uses C2PA metadata and invisible watermarking to help identify images made with its tools.
What is GPT Image?
GPT Image is the name OpenAI uses for its natively multimodal image models in the API. They take text and images as input and return images, and they power image generation in ChatGPT under the consumer name ChatGPT Images. Each release has had an API identifier (for example gpt-image-2) and, from version 2.5, two variants: Sunburst, which OpenAI describes as suited to workflows where editing precision matters most, and Flare, which it describes as suited to fast, high quality everyday generation.
Key facts
| Field | Value (as of 2026-11-15) |
|---|---|
| Provider | OpenAI |
| Modality | Input: text and image. Output: image |
| Current API models | gpt-image-2.6-sunburst, gpt-image-2.6-flare |
| Other active models | gpt-image-2.5-sunburst, gpt-image-2.5-flare (no deprecation listed) |
| Deprecated, still callable until shutdown | gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest, gpt-image-2 |
| First API release | gpt-image-1, 23 April 2025 |
| Endpoints | Image API (v1/images/generations, v1/images/edits) and the image generation tool in the Responses API |
| Open weights | No; API and ChatGPT only |
| Release dates for 2.6 | Not stated on OpenAI's model pages |
| Provenance | C2PA metadata and invisible watermarking (OpenAI, September 2026) |
Models and versions
The table lists each model OpenAI documents, oldest first, with the date evidence the official pages give. Where a page gives only a snapshot name, the table says so rather than treating it as a release date.
| Model | Date evidence | Status (as of 2026-11-15) |
|---|---|---|
| gpt-image-1 | Announced in the API on 23 April 2025 | Shut down 23 October 2026 |
| gpt-image-1-mini | Not stated | Deprecated; shutdown 1 December 2026 |
| gpt-image-1.5 | Snapshot gpt-image-1.5-2025-12-16; model page calls it "Our previous image generation model" | Deprecated; shutdown 1 December 2026 |
| chatgpt-image-latest | Not stated | Deprecated; shutdown 1 December 2026 |
| gpt-image-2 | ChatGPT Images 2.0 announced 21 April 2026; snapshot gpt-image-2-2026-04-21 | Deprecated; shutdown 10 February 2027 |
| gpt-image-2.5-sunburst, gpt-image-2.5-flare | ChatGPT Images 2.5 announced 8 September 2026; snapshot gpt-image-2.5-sunburst-2026-09-08 | Active |
| gpt-image-2.6-sunburst, gpt-image-2.6-flare | Release date not stated on the model pages | Active; listed replacement for all deprecated GPT Image models |
The ChatGPT Images 2.5 announcement says Flare is "the default choice for most applications" and gives higher quality than GPT-Image-2 at 50% lower latency, while Sunburst targets premium workflows that need tighter edit control and accepts longer generation times. The same post states that people create more than 3 billion images a week across ChatGPT Images and the GPT Image models in the API, a provider-reported figure.
Which GPT Image model should a team use?
The criteria below are factual, drawn from OpenAI's own descriptions. They are not a ranking.
- New integrations: OpenAI lists the 2.6 models as the replacement for every deprecated GPT Image model, so code written now against an older identifier will need changing again.
- Editing precision: OpenAI positions Sunburst for detailed visual work and precise edits, with longer generation times.
- Speed and volume: OpenAI positions Flare for fast iteration and everyday generation.
- Vector or animated output: SVG and animated GIF output, markup-guided edits and complexity control are documented for the 2.6 variants.
- Bulk, non-interactive jobs: the Batch API halves token prices for the 2.5 and 2.6 models.
For how GPT Image sits beside other image generators, with identical fields for each, see AI Image Generators Compared.
Pricing and plans
GPT Image is billed per token. Image output tokens dominate the cost, and the number of output tokens depends on the size and quality requested. Prices below are in US dollars per 1M tokens, from OpenAI's pricing and model pages as of 15 November 2026.
| Model | Text input | Cached text | Image input | Cached image | Image output | Batch image output |
|---|---|---|---|---|---|---|
| gpt-image-2.6-sunburst | $5.00 | $1.25 | $8.00 | $2.00 | $30.00 | $15.00 |
| gpt-image-2.6-flare | $5.00 | $1.25 | $8.00 | $2.00 | $30.00 | $15.00 |
| gpt-image-2.5-sunburst | $5.00 | $1.25 | $8.00 | $2.00 | $30.00 | $15.00 |
| gpt-image-1.5 (deprecated) | $5.00 | $1.25 | $8.00 | $2.00 | $32.00 | Not stated by OpenAI |
| gpt-image-1 at launch, April 2025 (shut down) | $5.00 | Not stated by OpenAI | $10.00 | Not stated by OpenAI | $40.00 | Not stated by OpenAI |
- Batch: OpenAI's Batch table lists the 2.5 and 2.6 models at half the standard rate: $2.50 text input, $4.00 image input, $15.00 image output, with cached inputs also halved.
- Worked example: the image generation guide gives an example output of 823 tokens costing $0.02469 at $30 per 1M image output tokens.
- Cached input: cached pricing applies only to the image generation tool in the Responses API, not to direct Image API requests.
- Streaming: each partial image adds 100 output tokens.
- Responses API: requests through the Responses tool also bill the mainline model's own token usage.
- Complexity: changing the 2.6 complexity setting carries no separate fee, though higher complexity can produce more output.
- History: at launch in April 2025, OpenAI put gpt-image-1 at roughly $0.02, $0.07 and $0.19 per square image for low, medium and high quality.
Features
Two ways to call it
- Image API: Generations creates an image from a text prompt; Edits modifies an existing image with a new prompt, partly or fully. OpenAI recommends this for a single image from one prompt.
- Responses API image tool: image generation as a built-in tool, with multi-turn editing and File IDs as inputs. The action parameter takes auto (default), generate or edit; forcing edit without an image in context returns an error.
Parameters and limits
| Setting | Documented values |
|---|---|
| size | Recommended 1024x1024, 1536x1024, 1024x1536, or auto. Sunburst, Flare and gpt-image-2 accept custom sizes: edges multiples of 16, aspect ratio between 1:3 and 3:1, no edge over 3840 px, total pixels 655,360 to 8,294,400. Above 2560x1440 is experimental |
| quality | Sunburst and Flare: low, medium, high, xhigh, max, auto (default auto). gpt-image-2: low, medium, high, auto. Earlier models up to high |
| complexity (2.6 only) | low, medium, high; default high when omitted |
| output_format | png (default), jpeg, webp; output_compression 0 to 100 for JPEG and WebP. 2.6 adds svg (generations, n=1, no streaming) and gif (frame_count 2 to 16, default 9; frame_duration_ms 20 to 2000, default 200) |
| background | transparent, with png or webp |
| n | Number of images per call; default 1 |
| partial_images | 0 to 3 when streaming |
| moderation | auto (default, standard filtering) or low (less restrictive) |
| input_fidelity | Omit for gpt-image-2, which always processes image inputs at high fidelity |
| mask and markup | Mask applies to the first image, same format and size, under 50 MB, with an alpha channel. Markup edits (2.6) use the first image as target; image and annotation layer each up to 16,777,216 pixels; do not combine markup with mask |
OpenAI lists these limitations: complex prompts can take up to 2 minutes, text rendering can still be imprecise, consistency of recurring characters or brand elements can vary, and precise placement in layout-sensitive compositions can be difficult. It also notes that masking is prompt-based and may not follow the mask's exact shape.
Availability
Yes. In ChatGPT the same family appears as ChatGPT Images. OpenAI said on 8 September 2026 that Images 2.5 was rolling out to ChatGPT, ChatGPT Work and Codex users across all tiers, on desktop, mobile and web. The announcement lists these ChatGPT features:
- Sketch: draw a visual guide in ChatGPT, invoked by typing @Sketch.
- Templates: starting formats such as Poster and Merch.
- Image comments: place comments on an image for focused edits.
- Prompt sharing: share the prompt behind an image so others can adapt it with their own photos.
Per-plan image limits in ChatGPT are not stated in the announcement. Organisations using ChatGPT Work or Enterprise should treat image generation as part of the workspace's data and retention settings; see ChatGPT Work Security.
Provenance and labelling
In its ChatGPT Images 2.5 announcement, OpenAI states that it uses C2PA metadata and invisible watermarking to help identify images made with its tools. The gpt-image-1 launch post in April 2025 said generated images include C2PA metadata. At Google I/O on 20 May 2026, Google said OpenAI was among the partners bringing Google's SynthID watermark to more AI-generated content; OpenAI's own September 2026 post does not name SynthID. C2PA metadata can be stripped when a file is re-encoded or screenshotted, so labelling rules should not rely on it alone. See SynthID vs C2PA and the AI Content Provenance and C2PA Guide.
Data and privacy
- Content filtering: prompts and images are filtered under OpenAI's content policy; blocked requests may return moderation_details with a moderation_stage of input, output or unknown.
- End-user identifiers: OpenAI recommends a safety_identifier, a privacy-preserving hash of 128 characters or fewer, for apps with end users; the user parameter is deprecated in its favour.
- Training on API data: the April 2025 launch post states OpenAI does not train on API customer data by default.
- System card: the 2.5 announcement links a system card for safety details; this page does not summarise it.
Where identity and security fit
Every GPT Image integration runs on an OpenAI API key or service account, and image endpoints are often added to an existing project rather than a dedicated one. Scope keys to the project that needs them, rotate them, and remove them when an integration is retired. The OpenAI API Keys Security Guide and How to audit OpenAI and Anthropic organisation API keys cover this in depth. Staff using ChatGPT Images through personal accounts for work content create shadow AI exposure; the AI Image and Video Generation at Work guide covers the policy side.
Release history
OpenAI released gpt-image-1 in the API on 23 April 2025, followed by gpt-image-1.5, gpt-image-2 (ChatGPT Images 2.0, 21 April 2026), the 2.5 Sunburst and Flare pair (8 September 2026) and the 2.6 pair, whose release date OpenAI's model pages do not state. Older models are retired on the dates in OpenAI's deprecations page.
When do the GPT Image models shut down?
OpenAI's deprecations page is the authoritative source. Its policy gives image generation models "at least 3 months" notice, unless safety or compliance concerns require a faster timeline; preview models can be retired with much shorter notice.
| Model (API identifier) | Notice date | Shutdown date | Listed replacement |
|---|---|---|---|
| dall-e-2 | 14 November 2025 | 12 May 2026 (past) | gpt-image-2, gpt-image-1 or gpt-image-1-mini |
| dall-e-3 | 14 November 2025 | 12 May 2026 (past) | gpt-image-2, gpt-image-1 or gpt-image-1-mini |
| gpt-image-1 | 22 April 2026 | 23 October 2026 (past) | gpt-image-2.5-sunburst or gpt-image-2.5-flare |
| gpt-image-1-mini | 2 June 2026 | 1 December 2026 | gpt-image-2.6-sunburst or gpt-image-2.6-flare |
| gpt-image-1.5, gpt-image-1.5-2025-12-16 | 2 June 2026 | 1 December 2026 | gpt-image-2.6-sunburst or gpt-image-2.6-flare |
| chatgpt-image-latest | 2 June 2026 | 1 December 2026 | gpt-image-2.6-sunburst or gpt-image-2.6-flare |
| gpt-image-2, gpt-image-2-2026-04-21 | 10 November 2026 | 10 February 2027 | gpt-image-2.6-sunburst or gpt-image-2.6-flare |
Two points catch teams out. First, some replacements listed at the time are themselves now retired: the DALL-E entries still name gpt-image-1 and gpt-image-1-mini, both of which have later shutdown dates of their own. Second, the gpt-image-2 alias and its dated snapshot share a single shutdown date, so pinning the snapshot does not buy extra time. Shutdown dates for the wider OpenAI catalogue, including Sora, are collected in the AI Media Model Retirements Tracker.
What happened to DALL-E?
DALL-E 2 and DALL-E 3 were OpenAI's earlier image models. OpenAI gave notice on 14 November 2025 and shut down dall-e-2 and dall-e-3 in the API on 12 May 2026. From early 2024, OpenAI attached C2PA metadata to images that DALL-E 3 created or edited in ChatGPT and the API, according to its May 2024 provenance post, and edits were recorded in the credentials. The same post described an internal classifier that identified about 98% of DALL-E 3 images in OpenAI's tests while wrongly flagging under 0.5% of non-AI images, offered to early testers through a researcher access programme. Those figures relate to DALL-E 3 only and are provider-reported.
How do you migrate off a retiring GPT Image model?
- Inventory callers: search code, configuration, infrastructure templates and low-code tools for dall-e-2, dall-e-3, gpt-image-1, gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest and gpt-image-2, including dated snapshots.
- Check parameters: quality values, size rules and input_fidelity differ by model; remove input_fidelity for gpt-image-2 and check that quality values such as xhigh and max are only sent to models that accept them.
- Re-estimate cost: token prices and token counts per image differ between generations. Older models used fixed token counts per quality (for a square image, 272 tokens at low, 1056 at medium and 4160 at high), while the newer models bill on actual output tokens.
- Test in a pilot: run representative prompts and edits against the replacement, compare outputs and latency, and check moderation behaviour, including moderation_details on blocked requests.
- Switch with a rollback path: make the model identifier a configuration value so it can be changed without a release, and keep the old identifier available until its shutdown date in case the pilot misses a regression.
- Close the old path: after shutdown, delete references, alerts and dashboards for the retired identifier.
Limits
- Verification: organisations may need to complete API Organization Verification before they can call GPT Image models.
- Short model lifetimes: gpt-image-1 was shut down 18 months after its API release, and every older GPT Image model now has a shutdown date, so integrations need a planned upgrade path.
- No open weights: the models are reached only through the API and ChatGPT, so there is no self-hosted option.
- Undated releases: OpenAI's model pages do not state a release date for the 2.6 models.
Practitioner checklist
- Search all repositories and configuration for retiring GPT Image and DALL-E identifiers, including dated snapshots.
- Record 1 December 2026 and 10 February 2027 as hard dates in the change calendar for the affected owners.
- Move model identifiers into configuration so a swap needs no code release.
- Pilot gpt-image-2.6-sunburst or gpt-image-2.6-flare with representative prompts and compare cost per image.
- Complete API Organization Verification ahead of migration if the organisation has not done so.
- Send a hashed safety_identifier for end-user apps and stop using the deprecated user parameter.
- Decide whether moderation stays at auto, and document any use of low.
- Keep C2PA metadata intact in publishing pipelines and add visible AI labels where disclosure rules require them.
- Revoke API keys used only by retired image integrations.
Standards and references
- OpenAI: Deprecations
- OpenAI: Image generation guide
- OpenAI: API pricing
- OpenAI: GPT-Image-2.6 Sunburst model page
- OpenAI: GPT-Image-2.6 Flare model page
- OpenAI: GPT-Image-2.5 Sunburst model page
- OpenAI: GPT-Image-2 model page
- OpenAI: GPT Image 1.5 model page
- OpenAI: Introducing ChatGPT Images 2.5
- OpenAI: Introducing ChatGPT Images 2.0
- OpenAI: Image generation API announcement (gpt-image-1)
- OpenAI: Understanding the source of what we see and hear online
- Google: I/O 2026 announcements
Related NHI Mgmt Group resources: AI Image, Video, Voice and Music Models Hub · OpenAI Models · OpenAI Sora Shutdown Explained · ChatGPT Enterprise Security Guide