OpenAI ships GPT Image 2.5 with two specialized models and multi-turn editing

A two-model image pipeline lands in OpenAI

OpenAI released GPT Image 2.5 with two variants: Flare for high-speed production and Sunburst for precision editing, shipped on September 8, 2026. A maintainer of a third-party wrapper notes that callers can select the new Flare and Sunburst variants without changing their existing default model, which lowers the integration cost for teams already routing image traffic through OpenAI or fal.

For practitioners, the split means model choice is now part of the brief — pick Flare when throughput dominates (catalog renders, batch mockups, social variants) and Sunburst when the edit must respect existing geometry or product identity.

Multi-turn editing: revise, don't restart

The new release centers on multi-turn editing: apply a series of AI photo edits to the same image step by step instead of re-prompting from scratch. OpenAI's own API guide recommends identifying what should change and what must stay the same, refining one thing at a time, and inspecting the result between turns — a discipline borrowed from non-destructive photo editing.

For product photography, that workflow is what makes iterative retouching viable. A blog walkthrough of product photo editing in GPT Image 2.5 demonstrates editing one part of an image while leaving the rest stable, which is the gap most previous generation tools left open.

Sketch tool pins the edit region

GPT Image 2.5 adds a sketch tool alongside the multi-turn flow, giving the model an explicit mask of where to act. Combined with the Sunburst variant, it gives creators a way to mark a region, request a change, and keep everything outside that region stable across turns. A social roundup highlights better detail and more consistent edits as the user-facing payoff.

The practical effect for a designer hand-off is that instructions can travel with the image: where to paint, where to keep, and which variant to use for which pass.

Higher-resolution output raises the export bar

GPT Image 2.5 raises output resolution, which shifts the export targets creators must design for. A Facebook creator thread testing ChatGPT image 2.5 for Instagram carousel creation reports that the new pipeline produces assets suitable for carousels without upscaling tricks, and the broader coverage flags improved image quality alongside the editing changes.

Higher source resolution makes downstream compression choices matter more. When shipping to the web, the format question — WebP versus JPG — becomes a real budget decision rather than an academic one, and the right pick depends on whether the source is photographic or flat-designed. Creators moving stills into a PDF deliverable also need a lossless path so the resolution gain survives the trip.

Practitioner workflow: which model when

The shape of the release — two models, sketch tool, multi-turn editing, higher-resolution output — suggests a layered pipeline rather than a single prompt. One route: draft with Flare for speed, switch to Sunburst for the final pass with sketch marks where the edit must land, then export at the new higher resolution for web and social. A lossless export step into PDF preserves the resolution for client deliverables, and the same source can be split for grid posting when the channel demands it.

Teams that already maintain a default model on OpenAI or fal can opt into the new variants without a migration, which makes a pilot test cheap.

What to verify next

Two things to check before locking the workflow in. First, confirm which Sunburst use cases actually need the sketch mask versus a plain prompt — wasted masks slow the turn loop. Second, confirm the new output resolution against the channels you ship to: a higher source asset changes both the compression budget and the size of any PDF hand-off. OpenAI's own API guide frames it as a "refine one thing at a time" discipline, and that same discipline now applies to choosing between Flare and Sunburst at the start of a job rather than midstream.