On this page · 12 sections
- What shuts down, and when
- What GPT-Image-2.6 Sunburst and Flare are
- Two gaps to check before you switch
- Price per token: what changes
- Price per image: the calculator changes the picture
- Worked monthly example
- Rate limits and access
- Azure OpenAI dates differ
- A two-week plan for the 1 December group
- Test checklist
- How eCorpIT can help
- References
Summary. OpenAI's deprecations page, checked on 17 November 2026, lists 1 December 2026 as the shutdown date for gpt-image-1.5 (snapshot gpt-image-1.5-2025-12-16), gpt-image-1-mini and chatgpt-image-latest, all announced on 2 June 2026. A second notice, dated 10 November 2026, retires gpt-image-2 and its gpt-image-2-2026-04-21 snapshot on 10 February 2027. For all five IDs OpenAI names the same replacements, gpt-image-2.6-sunburst or gpt-image-2.6-flare, released on 4 November 2026. That leaves 14 days for the first group and 85 days for gpt-image-2. GPT-Image-2.6 charges $30 per million image output tokens, against $32 for gpt-image-1.5 and $8 for gpt-image-1-mini. The lower rate does not mean a cheaper image: OpenAI's own calculator estimates a medium-quality square 2.6 image at about $0.043, against $0.034 on gpt-image-1.5 and $0.011 on mini. Two gaps matter: the 2.6 model pages show no Batch API support, and OpenAI documents 2.6 through the Image API only, while its Responses API examples still name GPT-Image-2.5. Our verdict: move direct Image API calls to Flare or Sunburst now, measure tokens per image, and hold Batch and Responses workloads on GPT-Image-2.5 until OpenAI confirms 2.6 support.
About this article. This is for engineering leads, product owners and finance teams who generate or edit images through OpenAI's API or Azure. We checked every model ID, date, price and limit on 17 November 2026 against OpenAI's deprecations page, model pages, pricing page, image generation guide and changelog, and against Microsoft's Foundry retirement schedule. Vendors change prices and dates often, so re-check before you commit; this is not legal or financial advice, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What shuts down, and when
OpenAI's table of upcoming deprecations carries two image-model notices, and both point to the same pair of replacement models.
| Model ID | Snapshot | Notice date | Shutdown date | Named replacement |
|---|---|---|---|---|
| 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 |
| gpt-image-1-mini | gpt-image-1-mini | 2 June 2026 | 1 December 2026 | gpt-image-2.6-sunburst or gpt-image-2.6-flare |
| chatgpt-image-latest | 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 |
Each of the four model pages already shows a Deprecated status. Three planning details follow.
- Alias and snapshot fail together. The
gpt-image-1.5alias resolves to the December 2025 snapshot andgpt-image-2to the April 2026 snapshot, and both snapshots are marked deprecated, so pinning the dated string buys no extra time.
- The notice periods follow OpenAI's policy. Image generation models get at least three months' notice. The 1 December group received 182 days; gpt-image-2 gets 92 days, just over the minimum.
- One model page disagrees with the table. The
chatgpt-image-latestpage still recommends GPT-Image-2.5 Sunburst for API use, while the deprecations table names the 2.6 pair. Treat the deprecations table as the current answer.
This is the third image-model shutdown date of 2026. DALL-E 2 and DALL-E 3 left the API on 12 May 2026, and gpt-image-1 shut down on 23 October 2026 with GPT-Image-2.5 as its named replacement. Teams that moved to gpt-image-2 rather than 2.5 now face a second migration by February. Our comparison of AI model retirement policies sets OpenAI's notice tiers against other vendors.
What GPT-Image-2.6 Sunburst and Flare are
OpenAI released both models on 4 November 2026 for image generation and editing, with adjustable complexity, SVG output and animated GIF output, and added markup-guided editing for supported GPT-Image-2.5 and 2.6 models. OpenAI positions Sunburst for detailed compositions and editing, and Flare for interactive, fast-iteration workflows. Both take text and image input and return images. Neither page lists a dated snapshot: the only model ID is the alias itself.
The new controls, from OpenAI's image generation guide:
- Complexity. A
complexityfield sets how much the model plans before rendering, withlow,mediumandhigh. The default is high. OpenAI says lower values suit focused edits such as changing a background, higher values suit dense text and complex layouts, and higher values can add latency and unwanted changes. Complexity carries no separate fee.
- Quality. The guide lists
low,medium,high,xhighandmaxfor both 2.6 models, againstlow,mediumandhighfor gpt-image-2.
- Markup-guided edits. On
POST /v1/images/edits, you send an annotation overlay that shows where the change should happen, plus a prompt. These edits use normal edit token rates with no extra fee.
- SVG output. On the generations endpoint only, with
nset to 1 and no streaming, partial images or output compression. The API first renders a reference image, then produces static vector content. You pay the image model's rate for the reference image and the authoring model's rate for the tokens used to write the SVG.
- Animated GIF output. On both generations and edits, with
frame_countfrom 2 to 16 (default 9) andframe_duration_msfrom 20 to 2,000 (default 200). No streaming or partial images.
A minimal request for a team moving a gpt-image-1.5 call to Flare looks like this. Keep quality and size explicit so that cost tests compare like with like.
{
"model": "gpt-image-2.6-flare",
"prompt": "Product photo of a steel water bottle on a plain grey background",
"size": "1024x1024",
"quality": "medium",
"complexity": "medium"
}
Send it to POST /v1/images/generations. If you omit complexity, the model runs at high, which may raise latency against your current baseline.
Two gaps to check before you switch
The deprecations table names 2.6 as the replacement for every retiring model, but the 2.6 pages do not yet match every way the old models were used.
Batch API. The gpt-image-1.5, gpt-image-1-mini, chatgpt-image-latest, gpt-image-2 and GPT-Image-2.5 pages all show v1/batch as supported. The GPT-Image-2.6 Sunburst and Flare pages show it as not supported. OpenAI's pricing page nevertheless lists Batch rates for 2.6 at half the standard price. Because OpenAI's own pages conflict, test a small batch job against 2.6 before you depend on it. If it fails, GPT-Image-2.5 Sunburst or Flare keeps Batch support at the same token rates and is not on the deprecation list as of 17 November.
Responses API image tool. OpenAI's guide says the 2.6 features work through the Image API, and its Responses API instructions still tell you to name a GPT-Image-2.5 model in the image generation tool's model field. The 4 November changelog entry lists only the generations and edits endpoints for 2.6, while the 8 September entry for 2.5 named both the Image API and the Responses tool. If your product calls image generation as a tool inside a Responses conversation, plan on GPT-Image-2.5 for now and recheck the guide before February.
Price per token: what changes
OpenAI bills these models per million tokens, split into text and image modalities. Standard rates on 17 November 2026, from OpenAI's pricing page and the individual model pages:
| Model | Text input | Cached text | Image input | Cached image | Image output |
|---|---|---|---|---|---|
| gpt-image-1.5 | $5.00 | $1.25 | $8.00 | $2.00 | $32.00 |
| chatgpt-image-latest | $5.00 | $1.25 | $8.00 | $2.00 | $32.00 |
| gpt-image-1-mini | $2.00 | $0.20 | $2.50 | $0.25 | $8.00 |
| gpt-image-2 | $5.00 | $1.25 | $8.00 | $2.00 | $30.00 |
| gpt-image-2.5 Sunburst or Flare | $5.00 | $1.25 | $8.00 | $2.00 | $30.00 |
| gpt-image-2.6 Sunburst or Flare | $5.00 | $1.25 | $8.00 | $2.00 | $30.00 |
gpt-image-1.5 and chatgpt-image-latest also bill text output at $10 per million; the newer models do not bill text output. On rates alone:
- gpt-image-1.5 and chatgpt-image-latest users see the image output rate fall from $32 to $30 per million tokens, about 6% lower. Input rates do not change.
- gpt-image-1-mini users pay 3.75 times more per image output token ($8 to $30), 2.5 times more for text input ($2 to $5) and 3.2 times more for image input ($2.50 to $8). OpenAI names no cheaper image model as a replacement.
- gpt-image-2 users see the same token rates on 2.6. Any cost change comes from token counts per image.
One more billing detail: for GPT-Image-2 and 2.5, cached input pricing applies only to the image generation tool in the Responses API, not to direct Image API calls such as edits. OpenAI does not say how this works for 2.6, so teams that move edits to the Image API should not assume cache discounts.
Price per image: the calculator changes the picture
OpenAI publishes per-image output prices for the retiring models.
| Model and quality | 1024x1024 | 1024x1536 | 1536x1024 |
|---|---|---|---|
| gpt-image-1.5, low | $0.009 | $0.013 | $0.013 |
| gpt-image-1.5, medium | $0.034 | $0.050 | $0.050 |
| gpt-image-1.5, high | $0.133 | $0.200 | $0.200 |
| gpt-image-1-mini, low | $0.005 | $0.006 | $0.006 |
| gpt-image-1-mini, medium | $0.011 | $0.015 | $0.015 |
| gpt-image-1-mini, high | $0.036 | $0.052 | $0.052 |
| gpt-image-2, low | $0.006 | $0.005 | $0.005 |
| gpt-image-2, medium | $0.053 | $0.041 | $0.041 |
| gpt-image-2, high | $0.211 | $0.165 | $0.165 |
The gpt-image-1.5 figures match OpenAI's token counts for models before gpt-image-2: 1,056 tokens for a medium square image at $32 per million is $0.0338, and 4,160 tokens for a high square image gives $0.133.
For 2.6, OpenAI gives no fixed table, but the cost calculator in its guide estimates output tokens per quality and size, and warns that models can use different token counts at the same quality setting. The estimates below are the calculator's, multiplied by $30 per million.
| GPT-Image-2.6 quality | 1024x1024 tokens | 1024x1024 cost | 1024x1536 or 1536x1024 cost |
|---|---|---|---|
| low | 823 | $0.025 | $0.020 |
| medium | 1,449 | $0.043 | $0.034 |
| high | 1,756 | $0.053 | $0.041 |
| xhigh | 3,122 | $0.094 | $0.074 |
| max | 7,024 | $0.211 | $0.165 |
Set against the old tables, a square 2.6 image at low or medium quality is estimated to cost more than gpt-image-1.5 at the same setting ($0.025 against $0.009, $0.043 against $0.034), while 2.6 at high is far cheaper ($0.053 against $0.133). Quality labels are not equivalent across models, so compare output you would actually ship. The calculator also estimates GPT-Image-2.5 at 439 tokens ($0.013) for a medium square image, a third of the 2.6 estimate. Each partial image in a streamed request adds 100 output tokens. Treat all of these as estimates and confirm them from the usage object in each response. Our earlier article on gpt-image-2's transparent backgrounds and missing token table covers how this calculator approach began with gpt-image-2.
Worked monthly example
Take a catalogue team producing 100,000 medium-quality square images a month, ignoring prompt and input-image tokens, which add to every row.
| Scenario | Arithmetic | Monthly output cost |
|---|---|---|
| gpt-image-1.5 today | 100,000 × $0.034 | $3,400 |
| gpt-image-1-mini today | 100,000 × $0.011 | $1,100 |
| gpt-image-2 today | 100,000 × $0.053 | $5,300 |
| GPT-Image-2.6 at 1,449 tokens | 100,000 × 1,449 × $30 per million | $4,347 |
| GPT-Image-2.5 at 439 tokens | 100,000 × 439 × $30 per million | $1,317 |
The 2.6 and 2.5 rows use the calculator estimates above; replace them with your measured average. At the ECB reference rates of 16 November 2026, crossed to ₹93.93 per US dollar and rounded to two decimals, $3,400 is about ₹3.19 lakh and $4,347 about ₹4.08 lakh a month. On these estimates a gpt-image-1.5 workload rises by about 28% and a mini workload roughly quadruples ($1,100 to $4,347). Our AI image generation API cost comparison and the Nano Banana shutdown guide cover Google's options if a second vendor makes sense for low-cost volume.
Rate limits and access
The retiring models publish the same limits: 20 images per minute and 250,000 tokens per minute at the Build tier, 150 and 3 million at Launch, and 250 and 8 million at Grow. The 2.6 pages publish no table and send you to your organisation's limits page instead. Check those limits before a cut-over, because a burst job sized for 250 images per minute may hit a lower ceiling.
OpenAI also says some organisations must complete API organisation verification before using GPT Image models. If you plan to send 2.6 traffic from a new project or organisation, verify early. For apps that serve end users, the guide recommends a hashed safety_identifier on each request; the older user parameter is deprecated but still accepted.
Azure OpenAI dates differ
Teams on Azure work to Microsoft's schedule, not OpenAI's. Microsoft's Foundry retirement page lists gpt-image-1.5 (version 2025-12-16) retiring on 16 December 2026, gpt-image-1-mini on 7 April 2027, the two GPT-Image-2.5 models on 9 September 2027 and gpt-image-2 on 21 October 2027, with no replacement named. GPT-Image-2.6 did not appear on that page when we checked. A team that uses Azure as a fallback keeps gpt-image-1.5 there for 15 days after OpenAI's shutdown, and gpt-image-2 for over eight months longer. That buys time, not a plan: the prompt and quality tests still have to run on whichever model you settle on.
If edits include customer photos or documents, record where image inputs are processed in your data-protection records alongside the model change.
A two-week plan for the 1 December group
Today is 17 November, so there are 14 days left.
| Dates | Work | Exit criterion |
|---|---|---|
| 17 to 18 Nov | Search code, configs, gateway routes and the usage dashboard for gpt-image-1.5, gpt-image-1-mini, chatgpt-image-latest and the dated snapshot; tag each call as Image API, Responses tool or Batch | Owner and path per call |
| 19 to 22 Nov | Build a test set of 100 to 200 real prompts and edits per workload; record tokens, latency and pass rate on the current model | Baseline saved |
| 23 to 26 Nov | Run the set on Flare and Sunburst at two complexity levels; run Batch and Responses-tool workloads on GPT-Image-2.5 as well | One target per workload |
| 27 to 29 Nov | Canary at 10% then 100% behind a feature flag; watch 429 errors and moderation blocks | Full traffic moved |
| 30 Nov | Block the retiring IDs in CI and alert on the shutdown_date field from the Models API | No old IDs in logs |
| 1 Dec | Shutdown | Requests to the three models fail |
OpenAI's deprecations page notes that the List models and Retrieve model endpoints return an optional shutdown_date field, which makes the alert in the last row easy to automate. Keep the test set for gpt-image-2's removal on 10 February 2027. Our GPT-5 and o3 shutdown plan covers the text models that leave on 11 December.
Test checklist
- Visual pass rate. A rubric or human score per prompt, with a regression threshold agreed in advance.
- Text in images. OpenAI still lists text placement and clarity among GPT Image limitations, so test labels, prices and signage separately.
- Edit fidelity. Compare masked edits against markup-guided edits on the same inputs, and check that untouched areas stay unchanged.
- Brand consistency. Recurring characters and brand elements across several generations, another listed limitation.
- Tokens and cost per accepted image. Input text, input image and output tokens from
usage, divided by images that pass review.
- Latency. OpenAI warns that complex prompts can take up to two minutes; record p50 and p95 at the complexity you plan to ship.
- Errors. Count
moderation_blockedresponses and other image generation user errors, which OpenAI says should not be retried without changing the input.
If a workload cannot move before 1 December, OpenAI's deprecations page says some developers may be able to provision dedicated capacity for continued access through its sales team. Treat that as an exception for contract-bound work, not a default.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our artificial intelligence and machine learning services team can inventory your image-model calls, run side-by-side tests on GPT-Image-2.6 and 2.5, and report cost per accepted image before you switch. Talk to us at /contact-us/.
Last updated: 17 November 2026.