GPT Image 2.5: Sunburst vs Flare, and the gpt-image-1 shutdown
GPT Image 2.5 Sunburst and Flare cost the same per token. What differs, what an image costs, and what to do before gpt-image-1 shuts down on 23 October.
GPT Image 2.5 arrived on 8 September as two models, Sunburst and Flare, and on OpenAI's own model pages they cost the same, share the same rate limits and accept the same quality settings. The only published difference is one sentence of positioning. By the end of this post you will know which one to call, what an image costs, and what to do before gpt-image-1 shuts down on 23 October.
- Released: 8 September 2026,
gpt-image-2.5-sunburstandgpt-image-2.5-flare, through the Image API and the Responses API image generation tool. - New quality settings:
xhighandmax, on both models. Earlier GPT Image models stop athigh. - Price: GPT Image 2 token rates, $30 per 1M image output tokens.
- Shutdowns:
gpt-image-1on 23 October 2026;gpt-image-1-mini,gpt-image-1.5andchatgpt-image-lateston 1 December 2026.
GPT Image 2.5 Sunburst vs Flare: what actually differs
OpenAI describes Sunburst as the choice "for workflows where editing precision matters most" and Flare as its "fastest model for high-quality, everyday image generation", in the image generation guide. That is the whole of the documented distinction. Put the two model pages side by side and the rest is identical, including support for the edits endpoint and inpainting on Flare.
| Published spec | Sunburst | Flare | gpt-image-2 |
|---|---|---|---|
| Positioning | Editing precision | Fast everyday generation | Previous model |
| Quality settings | low to max | low to max | low to high |
| Image output, per 1M tokens | $30 | $30 | $30 |
| Image input, per 1M tokens | $8 | $8 | $8 |
| Generations and edits endpoints | Yes | Yes | Yes |
| Batch API | Not supported | Not supported | Supported, 50% off |
The prices and endpoint support come from the Sunburst and Flare model pages; the gpt-image-2 Batch discount comes from its 21 April release note. OpenAI publishes no speed figure for Flare, so "fastest" is a claim you should time on your own prompts. Equal token rates also do not mean equal bills: the guide warns that "token consumption can differ by model and quality setting".
GPT Image 2.5 Sunburst vs Flare: which to use
Use Flare where a person is waiting for a first image: a generator inside your app, a draft for a marketing team. Use Sunburst where the job is changing an existing picture without breaking it, such as swapping a background behind a product shot or editing a masked region. If your workload is both, the cheapest test is to run the same 50 prompts through each at the same quality setting and read usage from the responses.
When is gpt-image-1 being shut down? 23 October, then three more on 1 December
OpenAI's deprecations page lists two separate waves, and both name gpt-image-2 as the replacement, not GPT Image 2.5.
| Shutdown date | Model | Recommended replacement |
|---|---|---|
| 23 October 2026 | gpt-image-1 | gpt-image-2 |
| 1 December 2026 | gpt-image-1-mini | gpt-image-2 |
| 1 December 2026 | gpt-image-1.5 | gpt-image-2 |
| 1 December 2026 | chatgpt-image-latest | gpt-image-2 |
The two OpenAI pages pull in different directions. The deprecations table points to gpt-image-2, while the image guide says "for new integrations, use one of the GPT Image 2.5 models". Both are live and priced identically per token, so the choice is yours. We cover the rest of the 23 October wave, including gpt-4 and o1, in our GPT-4 shutdown migration map.
How much does OpenAI image generation cost per image?
OpenAI image generation is billed per token, and the pricing page lists these rates per 1M tokens as of September 2026:
| Model | Image input | Image output | Text input |
|---|---|---|---|
| gpt-image-2.5-sunburst / flare | $8.00 | $30.00 | $5.00 |
| gpt-image-2 | $8.00 | $30.00 | $5.00 |
| gpt-image-1.5 | $8.00 | $32.00 | $5.00 |
| gpt-image-1 | $10.00 | $40.00 | $5.00 |
| gpt-image-1-mini | $2.50 | $8.00 | $2.00 |
Per-image cost is output tokens times the rate. The guide's older-model table shows a 1024x1024 image at high quality uses 4,160 output tokens on the earlier GPT Image models. On gpt-image-1 that is 4,160 x $40 / 1,000,000 = $0.166, which matches the $0.167 OpenAI lists. OpenAI does not publish token counts for GPT Image 2.5, and says its GPT Image 2 calculator "does not estimate GPT Image 2.5 token consumption". So the honest per-image figure for Sunburst or Flare is whatever your usage field reports, at $0.03 per 1,000 output tokens.
Two costs catch teams out. Streaming with partial_images adds 100 image output tokens per partial image, and gpt-image-2 always processes input images at high fidelity, so edit requests with reference images can use more input tokens. gpt-image-1-mini is also the cheapest model on the list, and its named replacement costs more per output token than it does.
Migrate gpt-image-1 to gpt-image-2 or straight to GPT Image 2.5?
Decide on one question first: do you generate images in bulk off the critical path? If you do, gpt-image-2 is the only one of the three with Batch support, at half price. If a user is waiting, GPT Image 2.5 gives you the extra quality settings. Then work through the changes:
- Swap the model ID in every config, environment variable and test fixture that names
gpt-image-1,gpt-image-1-mini,gpt-image-1.5orchatgpt-image-latest. - Drop
input_fidelityif you move togpt-image-2; the guide says to omit it because the model always uses high fidelity. - Check organisation verification. The guide notes you may need API Organization Verification before using GPT Image models.
- Re-measure cost on 20 to 50 real prompts per quality setting and compare the token counts, not just the rates.
- Keep a fallback behind a flag until the new model has served real traffic for a week.
gpt-image-2 transparent background support
Transparent backgrounds reached gpt-image-2 in preview on 20 August: set background to transparent and use png or webp output, because jpeg does not support transparency. The guide gives the same instruction for both GPT Image 2.5 models. For product imagery that goes onto coloured page backgrounds, this removes a background-removal step you may be running today.
What we would run for product imagery
In the e-commerce and marketplace builds we run, image generation usually splits into two jobs: a nightly batch of catalogue images and an interactive editor that a merchant uses in the admin panel. On current docs we would put the batch on gpt-image-2 through the Batch API and the editor on Sunburst, then test Flare for customer-facing previews. If that editor accepts customer photo uploads, the HEIC image upload vulnerability is a reminder to decode untrusted images carefully before any of them reach a model. We have not benchmarked Flare's speed against Sunburst's, and neither has OpenAI in public, so treat that split as a starting hypothesis to measure.
This is the wrong split if your volumes are small; at a few hundred images a month the Batch discount is not worth a second code path. If you want help wiring image generation into an e-commerce platform or a wider generative AI feature, that is work we do. The date to put in the calendar is 23 October: any request still naming gpt-image-1 after that stops working.
Frequently asked questions
GPT Image 2.5 Sunburst is positioned for workflows where editing precision matters most, and Flare for fast, high-quality everyday generation. OpenAI's model pages list identical token prices, identical rate limits, the same quality settings up to max, and edit endpoint support on both, so test each on your own prompts.
OpenAI shuts down gpt-image-1 on 23 October 2026 and recommends gpt-image-2 as the replacement. A second wave follows on 1 December 2026, when gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest are removed from the API, also with gpt-image-2 named as the replacement on the deprecations page.
OpenAI image generation is billed per token. GPT Image 2.5 and gpt-image-2 charge $30 per million image output tokens and $8 per million image input tokens. OpenAI publishes no token counts for GPT Image 2.5, so measure the usage field on real prompts to get a per-image cost for your sizes and quality settings.
GPT Image 2.5 Sunburst and Flare are listed as not supported on the Batch endpoint on their model pages as of September 2026. gpt-image-2 does support Batch, with a 50% discount, which makes it the cheaper option for bulk catalogue generation that does not need an immediate result.
gpt-image-2 supports transparent backgrounds in preview since 20 August 2026. Set background to transparent and request png or webp output, because jpeg does not support transparency. OpenAI's image generation guide gives the same instruction for both GPT Image 2.5 models, Sunburst and Flare.
