On this page · 11 sections
- Two deadlines, one replacement pair
- What changes when you move to 2.6
- Step 1: find every call before 1 December
- Step 2: sort each call by how it reaches OpenAI
- Step 3: change the request, and set every parameter
- Step 4: re-baseline cost per image
- Step 5: roll out in two waves
- Azure runs on a different calendar
- India-specific considerations
- How eCorpIT can help
- References
Summary. OpenAI's deprecations page, checked on 18 November 2026, lists two shutdowns for its image models. On 1 December 2026, 13 days away, gpt-image-1.5 (snapshot gpt-image-1.5-2025-12-16), gpt-image-1-mini and chatgpt-image-latest stop accepting API requests, under a notice dated 2 June 2026. On 10 February 2027, 84 days away, gpt-image-2 and its gpt-image-2-2026-04-21 snapshot follow, under a notice dated 10 November 2026. 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. The token rate barely moves: GPT-Image-2.6 charges $30 per million image output tokens, the same as gpt-image-2 and less than the $32 on gpt-image-1.5. But both 2.6 model pages mark the Batch API as not supported, while gpt-image-2 supports Batch at half price, and OpenAI's guide documents the Responses API image tool only with GPT-Image-2.5. OpenAI's own calculator also puts a medium 2.6 image at about four times the cost of a gpt-image-1-mini one. Our verdict: move direct Image API calls to 2.6 before 1 December, hold Batch and Responses workloads on GPT-Image-2.5, and measure cost per image rather than trusting the unchanged rate.
About this article. This is for engineering leads, platform teams and finance owners who call OpenAI's image models directly or through Azure. We checked every model ID, date, price and endpoint on 18 November 2026 against OpenAI's deprecations page, model pages, pricing page, image generation guide and cost calculator, API reference and changelog, and against Microsoft Foundry's retirement schedule and lifecycle policy. Rupee figures use European Central Bank reference rates. Vendors change dates and prices, 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.
Two deadlines, one replacement pair
OpenAI's deprecations page lists these image model shutdowns on 18 November 2026.
| Model ID | Snapshot covered | Notice date | Shutdown date | Days left on 18 Nov | Named replacement |
|---|---|---|---|---|---|
| gpt-image-1.5 | gpt-image-1.5-2025-12-16 | 2 Jun 2026 | 1 Dec 2026 | 13 | gpt-image-2.6-sunburst or flare |
| gpt-image-1-mini | none listed | 2 Jun 2026 | 1 Dec 2026 | 13 | gpt-image-2.6-sunburst or flare |
| chatgpt-image-latest | none listed | 2 Jun 2026 | 1 Dec 2026 | 13 | gpt-image-2.6-sunburst or flare |
| gpt-image-2 | gpt-image-2-2026-04-21 | 10 Nov 2026 | 10 Feb 2027 | 84 | gpt-image-2.6-sunburst or flare |
Four details shape the plan.
Snapshots fall with their alias. The deprecations table names both the alias and the dated snapshot for gpt-image-1.5 and gpt-image-2, and the gpt-image-2 model page marks its snapshot as deprecated. Pinning the dated string buys no extra time.
The gpt-image-2 notice is close to the minimum. OpenAI's policy promises image generation models at least three months' notice. The first wave received 182 days. gpt-image-2 gets 92, from 10 November 2026 to 10 February 2027.
Shutdown means requests fail. OpenAI defines a shut-down model as one that no longer accepts requests through the standard API. Nothing redirects to the replacement, so a hard-coded model string becomes an error on the day. The same page says some developers may be able to buy dedicated capacity for access after shutdown through OpenAI's sales team; treat that as a fallback to ask about, not a plan.
Some teams are on their third move of 2026. OpenAI removed DALL-E 2 and DALL-E 3 on 12 May 2026 and recommended gpt-image-2, gpt-image-1 or gpt-image-1-mini in their place. gpt-image-1 then shut down on 23 October 2026 with GPT-Image-2.5 as its replacement. Our first-wave briefing on the gpt-image-1.5 shutdown covers the 2.6 features in more depth; this is the runbook for both moves.
What changes when you move to 2.6
GPT-Image-2.6 Sunburst is positioned for detailed compositions and editing, and Flare for interactive, fast iteration. Both take text and image input and return images, and each lists only its alias, with no dated snapshot to pin. OpenAI's changelog says the release added adjustable complexity, SVG output and animated GIF output, plus markup-guided edits for supported 2.5 and 2.6 models.
For a migration, the request surface matters more than the feature list. The table draws on the three model pages, the changelog and the image generation guide.
| Capability | gpt-image-2 | GPT-Image-2.5 Sunburst | GPT-Image-2.6 Sunburst or Flare |
|---|---|---|---|
| Image generations and edits | Supported | Supported | Supported |
| Batch API | Supported | Supported | Not supported on model page |
| Responses API image tool | Yes, per changelog | Named in the guide | Not documented |
| Quality settings | low, medium, high | up to xhigh and max | up to xhigh and max |
| Complexity control | No | No | low, medium, high (default high) |
| SVG and GIF output | No | No | Yes |
| Deprecation status | Deprecated, 10 Feb 2027 | Not listed | Current |
Three rows need explanation.
Batch. The gpt-image-2 and GPT-Image-2.5 Sunburst pages list v1/batch as supported, and OpenAI's April release note for gpt-image-2 announced Batch at a 50% discount. The 2.6 Sunburst and Flare pages show v1/batch as not supported. OpenAI's pricing page nevertheless lists Batch rates for 2.6 at half the standard price, $15 per million image output tokens. When two OpenAI pages disagree, test before you depend on either. A small Batch job against 2.6 will settle the question for your account.
Responses API. Every image model page here, including gpt-image-2 and 2.5, shows v1/responses as not supported, because these models run inside the Responses API as the image generation tool rather than as the top-level model. What matters is the guide: it tells Responses API users to name gpt-image-2.5-sunburst or gpt-image-2.5-flare in the tool's model field, and says the 2.6 features work through the Image API. gpt-image-2 gained transparent backgrounds in the Responses tool in August 2026, so some conversational products use it there; their documented target is 2.5.
Complexity defaults to high. If you omit the new complexity field, both 2.6 models plan at high. The guide warns that higher complexity can add latency and introduce unwanted changes, and says complexity carries no separate fee. A like-for-like test against gpt-image-2 should set complexity explicitly, or the new model will look slower than it needs to.
Editing pipelines need one more check: gpt-image-2 always processes input images at high fidelity, which the guide says can raise input tokens on edits. Measure edits separately from generations.
Step 1: find every call before 1 December
Search code, configuration, infrastructure templates and any no-code tools for these strings: gpt-image-1.5, gpt-image-1-mini, chatgpt-image-latest and gpt-image-2. A search for gpt-image-2 also matches the 2.5 and 2.6 IDs, which shows where newer models are already in use.
Do not rely on code search alone. Model names often live in environment variables, feature flags, database rows that hold per-tenant settings, or a gateway's routing table. Your OpenAI usage dashboard, filtered by model, shows what is actually being called.
OpenAI also exposes shutdown dates through the API. The List models and Retrieve a model endpoints return an optional shutdown_date field, null when no date has been announced. A small scheduled check turns that into an alert.
from datetime import date
from openai import OpenAI
client = OpenAI()
WATCH = {"gpt-image-1.5", "gpt-image-1-mini", "chatgpt-image-latest",
"gpt-image-2", "gpt-image-2.5-sunburst", "gpt-image-2.5-flare",
"gpt-image-2.6-sunburst", "gpt-image-2.6-flare"}
for m in client.models.list():
if m.id in WATCH:
sd = getattr(m, "shutdown_date", None)
if sd:
days = (date.fromisoformat(str(sd)) - date.today()).days
print(f"{m.id}: shuts down {sd} ({days} days)")
else:
print(f"{m.id}: no shutdown date announced")
Run it daily in CI or a cron job and fail the build when any model you call is under 60 days from shutdown.
Step 2: sort each call by how it reaches OpenAI
The right target depends on the path each request takes. The table follows OpenAI's model pages and image guide as of 18 November 2026.
| Current path | Retiring model | Target today | Why |
|---|---|---|---|
| Image API, synchronous | Any of the four | gpt-image-2.6-sunburst or flare | Named replacement, fully documented |
| Batch API | Any of the four | GPT-Image-2.5 Sunburst or Flare | 2.6 model pages show no Batch support |
| Responses API image tool | gpt-image-2 | GPT-Image-2.5 Sunburst or Flare | Guide names 2.5 for the tool |
| Azure OpenAI or Foundry | Any | Follow Microsoft's schedule | Separate dates and replacements |
GPT-Image-2.5 is not on the deprecation list today, and its Sunburst page offers a dated snapshot, gpt-image-2.5-sunburst-2026-09-08, which 2.6 does not. The trade-off is that 2.5 is one generation behind and will get its own notice in time. Treat it as a bridge for the paths 2.6 does not cover, and recheck the 2.6 pages monthly.
Choosing between Sunburst and Flare is a quality and speed question, not a price question: both 2.6 models list identical token rates. Use Flare where a person waits for the image, and Sunburst for final assets and precise edits. If a gateway such as LiteLLM or Cloudflare AI Gateway sits in front of OpenAI, make the model a routing alias rather than a string in application code; our AI gateway comparison covers the options.
Step 3: change the request, and set every parameter
A minimal move from gpt-image-2 to Flare on the generations endpoint looks like this. Set size, quality and complexity explicitly, and read the usage object so cost is measured.
import base64
from openai import OpenAI
client = OpenAI()
result = client.images.generate(
model="gpt-image-2.6-flare", # was "gpt-image-2"
prompt="Product shot of a steel water bottle on a white desk, soft daylight",
size="1024x1024",
quality="medium",
output_format="jpeg",
extra_body={"complexity": "medium"}, # 2.6 only; omitted means high
)
with open("bottle.jpg", "wb") as f:
f.write(base64.b64decode(result.data[0].b64_json))
u = result.usage
print(u.input_tokens, u.output_tokens, u.total_tokens)
OpenAI's API reference lists complexity as an optional field with low, medium and high values, and documents a usage object on image responses with input and output token counts. If your SDK version does not yet type the field, pass it through extra_body as shown. Three parameter checks catch most regressions:
- Quality. gpt-image-2 accepts low, medium and high; 2.5 and 2.6 add xhigh and max. Keep your current value for the first comparison.
- Size. Sunburst and Flare accept custom sizes with edges in multiples of 16, an aspect ratio between 1:3 and 3:1, no edge above 3,840 pixels and a total between 655,360 and 8,294,400 pixels. Those limits match gpt-image-2's, so most size strings carry over.
- Input fidelity. The reference documents input_fidelity for gpt-image-1.5 and gpt-image-1-mini and says to omit it for gpt-image-2; it says nothing about 2.6. Remove it, or test before relying on it.
For the first wave, note that gpt-image-1.5 and chatgpt-image-latest bill text output at $10 per million tokens and image output at $32, while gpt-image-1-mini bills image output at $8; 2.6 bills no text output. Note that the chatgpt-image-latest model page still recommends GPT-Image-2.5 Sunburst for API use, while the deprecations table names 2.6.
Step 4: re-baseline cost per image
OpenAI's guide warns that equal token rates do not mean equal cost per image, and points to the usage object for measurement. The guide publishes a fixed per-image price table for gpt-image-2 and the older models; for 2.5 and 2.6 it offers only a calculator, which estimates output tokens from model, quality and size.
The figures check out. A medium 1024 by 1024 image on the pre-gpt-image-2 models uses 1,056 output tokens; at $32 per million that is $0.0338, which OpenAI rounds to $0.034 for gpt-image-1.5. The calculator puts gpt-image-2 at 1,756 tokens for the same image (1,756 × $30 per million = $0.0527, the published $0.053). For 2.6 it estimates 1,449 tokens at medium ($0.0435) and 823 at low ($0.0247); for 2.5 it estimates 439 at medium ($0.0132).
Here is what 100,000 medium square images a month cost in image output alone. The first four rows use OpenAI's published per-image prices; the 2.6 and 2.5 rows use the calculator estimates. Rupees use ECB reference rates of 17 November 2026, crossed to ₹93.48 per US dollar and rounded to two decimals.
| Model and mode | Output cost per image | Monthly cost, USD | Monthly cost, INR |
|---|---|---|---|
| gpt-image-1-mini, medium | $0.011 | $1,100 | ₹1,02,828 |
| gpt-image-1.5, medium | $0.034 | $3,400 | ₹3,17,832 |
| gpt-image-2, medium | $0.053 | $5,300 | ₹4,95,444 |
| gpt-image-2, medium via Batch | $0.0265 | $2,650 | ₹2,47,722 |
| GPT-Image-2.6, medium (estimate) | $0.0435 | $4,347 | ₹4,06,358 |
| GPT-Image-2.6, low (estimate) | $0.0247 | $2,469 | ₹2,30,802 |
| GPT-Image-2.5, medium (estimate) | $0.0132 | $1,317 | ₹1,23,113 |
The workings: 100,000 × $0.011 = $1,100, × $0.034 = $3,400 and × $0.053 = $5,300; the Batch line halves $5,300 because Batch image output is $15 against $30 per million tokens. The estimate rows are 100,000 × 1,449, 823 and 439 tokens × $30 per million.
Four conclusions follow.
- gpt-image-1-mini users face the largest change. On the calculator's estimate, 2.6 at medium costs about four times as much ($4,347 against $1,100), and even 2.6 at low costs more than twice as much ($2,469). GPT-Image-2.5 at medium is the closest in cost, at an estimated $1,317. Quality labels are not equivalent across models, so compare images you would actually ship.
- gpt-image-1.5 users pay about 28% more at medium on 2.6 ($3,400 to $4,347), despite the lower token rate.
- Batch users lose their discount if they move to 2.6 and the model page is right. A workload that paid $2,650 a month through Batch on gpt-image-2 would pay an estimated $4,347 on synchronous 2.6 at medium. GPT-Image-2.5 keeps Batch at $15 per million output tokens.
- Input tokens are extra. Prompts bill at $5 per million text input tokens and reference images at $8 per million image input tokens on 2.6. Edit-heavy pipelines should measure both.
Replace the estimates with averages from your own usage objects. For cross-vendor context, our AI image generation API cost comparison sets OpenAI against Google and others.
Step 5: roll out in two waves
- By 21 November. Finish the inventory, add the shutdown_date check to CI and agree on a target for each path. Pick 50 to 200 real prompts, including edits, as a golden set.
- By 25 November. Run the golden set on 2.6 Flare and Sunburst at your current quality, with complexity at low, medium and high. Record output tokens, latency and a human pass or fail per image.
- By 27 November. Ship the first-wave change behind a flag, route 10% of traffic and watch error rates and spend for 24 hours, then move to 100%. Keep at least two working days before 1 December for rollback.
- December. Repeat the golden set for gpt-image-2 workloads. Test a small Batch job on 2.6; if it fails, move Batch to 2.5.
- By 27 January 2027. Finish the gpt-image-2 cut-over, two weeks before the 10 February shutdown, and delete the old model strings so nothing falls back to them.
The same pattern works for OpenAI's text model retirements. Our GPT-5 and o3 shutdown migration guide applies it to 11 December 2026.
Azure runs on a different calendar
Teams on Azure OpenAI in Microsoft Foundry should not use OpenAI's dates. Microsoft's retirement schedule, dated 21 September 2026, lists different dates for the same models.
| Model | OpenAI API shutdown | Azure Foundry retirement |
|---|---|---|
| gpt-image-1.5 (2025-12-16) | 1 Dec 2026 | 16 Dec 2026 |
| gpt-image-1-mini (2025-10-06) | 1 Dec 2026 | 7 Apr 2027 |
| gpt-image-2 (2026-04-21) | 10 Feb 2027 | 21 Oct 2027 |
| GPT-Image-2.5 Sunburst and Flare (2026-09-09) | Not scheduled | 9 Sep 2027 |
Microsoft's schedule did not list GPT-Image-2.6 when we checked. Under Microsoft's lifecycle policy, a retired model returns 410 Gone to every inference request, and a deprecated model stays available to existing subscriptions but not to new ones. Azure customers on gpt-image-1.5 have 15 extra days; on gpt-image-2 they have until October 2027, but the newest models reach Azure later. Microsoft updates the schedule independently, so recheck it before you plan.
India-specific considerations
Indian teams pay OpenAI in US dollars, so every figure above moves with the exchange rate. At ₹93.48 per dollar, a $4,347 monthly image bill is ₹4,06,358 before any GST treatment your finance team applies to imported services. Budget the 2027 line on measured 2.6 tokens, not on the gpt-image-1-mini figures many teams started with.
Image prompts and edit inputs can carry personal data: customer photos, ID documents in KYC flows, or staff images in HR tools. While re-testing, check that the images you send to OpenAI are covered by your notice and purpose under the Digital Personal Data Protection Act 2023, and that you hash end-user identifiers. OpenAI's guide recommends a privacy-preserving safety_identifier of at most 128 characters, and says the Images API user parameter is deprecated in its favour.
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 AI and machine learning services team can inventory your OpenAI and Azure image calls, run golden-set tests on GPT-Image-2.6 and 2.5, and ship each change behind a flag before the deadline. Talk to us at /contact-us/.
Last updated: 18 November 2026.