As reviewed on 28 October 2026, ChatCut’s inspected integration exposes GPT Image 2, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. All three have small development-generation successes and a 10-reference ChatCut cap. Image 2 transparent output was explicitly rejected by the checked endpoint; Image 2.5 exposes a background control, but transparent output was not completed in this audit. Deployed production remains unverified.
Exact models and evidence
The selectors map to the same public model IDs shown here. Official specifications were accessed on 27 October 2026; pinned integration and saved endpoint checks were reviewed on 28 October. “Exposed” describes a control in that inspected integration, not a production entitlement guarantee.
| Model and public ID | Capabilities, integration and evidence |
|---|---|
GPT Image 2gpt-image-2Availability: Pinned integration and saved dev checks; production unverified. Source links describe native specifications. Reviewed: | Native documentation: Text generation and image edits; direct API documents up to 16 references and a 32,000-character prompt ceiling. ChatCut integration: Exact selector exposed; 10 combined references; no background control for Image 2. Endpoint evidence: One low-quality 1024×640 PNG succeeded. Transparent background explicitly rejected; high-resolution boundary outputs unverified. |
GPT Image 2.5 Flaregpt-image-2.5-flareAvailability: Pinned integration and saved dev checks; production unverified. Source links describe native specifications. Reviewed: | Native documentation: Generation and edits; documented quality and transparent-output controls; custom sizes follow pixel/edge rules. ChatCut integration: Exact selector, 10 combined references and Image 2.5 background control exposed. Endpoint evidence: One low-quality 1024×640 PNG succeeded. Transparent output and high-resolution boundaries unverified in endpoint audit. |
GPT Image 2.5 Sunburstgpt-image-2.5-sunburstAvailability: Pinned integration and saved dev checks; production unverified. Source links describe native specifications. Reviewed: | Native documentation: Generation and edits; documented quality and transparent-output controls; custom sizes follow pixel/edge rules. ChatCut integration: Exact selector, 10 combined references and Image 2.5 background control exposed. Endpoint evidence: One low-quality 1024×640 PNG succeeded. Transparent output and high-resolution boundaries unverified in endpoint audit; separate local scene run produced larger stills. |
Flare or Sunburst: which should I choose?
OpenAI positions Flare around speed and Sunburst around precise generation/editing. This review contains no controlled Flare-versus-Sunburst quality or timing comparison. Both document image editing and the same families of output controls. Choose an available model, make one small representative draft, and inspect product details, text, composition and the quoted cost before increasing quality. The local Sunburst example below demonstrates one workflow; it does not establish superiority.
References, edits and output options
For ChatCut, use at most 10 combined image references, counting project images and public-URL images together. Local validation accepted 10 and rejected 11. The direct OpenAI edit API documents 16; the saved 16/17 probes used invalid images and failed decoding, so they do not establish native boundary generation. Prepare PNG, JPEG or WebP; the native edit limit is less than 50 MB per image, while ChatCut’s public-URL path has its own 5 MiB limit. Rasterize SVG yourself. See generation references.
Keep prompts at 32,000 characters or fewer. All three checked endpoints rejected 32,001 characters, even though the inspected tool did not pre-check that maximum. Text-only generation needs no image. For an edit, reference the latest image and state what must remain fixed. OpenAI’s optional mask API is not proof that a mask control is exposed in ChatCut, and edits do not guarantee exact boundaries or identity preservation. Edit API.
ChatCut offers 1K/2K/4K labels that map to pixel budgets, not guaranteed 2048- or 4096-pixel edges. The checked 2.5 specification requires dimensions divisible by 16, each edge at most 3840, 655,360–8,294,400 total pixels and a 1:3–3:1 ratio. OpenAI marks dimensions above 2560×1440 experimental; this is not a blanket rule for every provider’s 2K/4K tier. Malformed-reference probes did not produce high-resolution outputs. A separate local scene run did produce larger Sunburst images, as scoped below. Image generation guide.
For backgrounds, do not request transparency with Image 2 on this route: the endpoint explicitly rejected it. The inspected integration exposes Auto/Opaque/Transparent only for Image 2.5. Its documented native transparency support and visible control do not prove a completed transparent result here. Inspect the actual alpha channel before relying on a cutout.
Reproduce a reference-led still workflow
Prepare one draft
Attach one clear product reference or use @ to reference a project image. Ask for one 16:9 image at 1K with an opaque background, and name the available GPT model explicitly.
Describe the edit
For example: keep the product silhouette and colors; place it on a plain gray surface with soft side light and empty space on the right; add no text. Keep the prompt within 32,000 characters.
Check and place
Review the model, settings and credit estimate if shown. Inspect the returned image and its dimensions, then drag it from My Assets to a video track. Adjust crop, scale and duration; reference that result for a later revision.
The result is a raster asset with editable timeline placement, crop and duration. Generated people, text and objects do not become separate editable layers. For the complete task tutorial, use image generation.
What did the local Sunburst run cost?
A 27 October 2026 local scene test, reviewed on 28 October, used gpt-image-2.5-sunburst, high quality and requested 16:9. It completed 20 casting sheets at requested 2K (actual 2304×1296) and 10 office plates at requested 4K (actual 3776×2128). Settled image charges were 4.00 credits for each group, 8.00 total. These are historical local-test catalog results; production cohort/catalog equivalence and current quotes were not verified. They are not a tariff or a Flare comparison.
The example has no video input or output duration. It reports settled totals, not reusable estimates or a verified rounding rule. The wider scene run also used video generation and 22.24 Agent credits separately; 8.00 is only its still-image spend. Before a new request, compare current quotes at matching model, count, quality, ratio, dimensions and reference settings, then inspect Credits history → Detail. Credits Policy owns current rates; generation credits explains reconciliation.
Frequently asked questions
Which aliases are Flare and Sunburst?
Is Sunburst better than Flare?
Can I submit 16 reference images through ChatCut?
Which transparency controls worked on the checked route?
Are the local Sunburst charges current public prices?
Last verified: 28 October 2026. Public sources: Image 2, Flare, Sunburst, edit API and generation guide, accessed 27 October. Endpoint results and the separately scoped local scene ledger were reviewed 28 October; no new paid generation was performed for this reference. For missing options or rejected requests, see workflow failures and the model matrix.