OpenAI ships GPT Image 2.5 with Flare and Sunburst API models

Two models, two jobs in the same image API

OpenAI introduced GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst side by side, splitting the product into a speed-oriented variant and a precision-oriented one. Flare carries the same quality, editing, and speed improvements, while Sunburst is positioned as the reference-aware model. Together they give API users a choice between latency and fidelity without leaving the same surface. For practitioners shipping image features, the practical move is to evaluate both on real reference sets before committing to one.

Faster generation and tighter edit control hit production

Speed is the headline workflow change. GPT Image 2.5 generates images up to 50% faster, and editing is described as making more precise edits while preserving previous changes across turns. The framing in public explainers calls this out as the upgrade from prior versions: edits only what was asked, with prior layers of change retained. For shippers, that closes one of the most common gaps in consumer-facing image tools — losing an earlier fix when the user makes a follow-up edit.

Reference images and resolution tiers change how inputs are packaged

Sunburst accepts natural-language instructions applied to up to 16 reference images, with an optional mask. Output is arbitrary resolutions up to 3840x2160, spanning five quality tiers including xhigh and max. Creators who used to pre-compose references in Photoshop or Canva now have an API path that takes the bundle directly, which changes how teams package brand kits and product shots before sending them in.

Consistency is the new selling point

OpenAI's official guide frames GPT Image 2.5 around precise editing and subject consistency, and the marketing around the launch uses the phrase "Power of Consistency." Independent commentary echoes the same point, listing reference-photo fidelity, precise edits, Sketch, Templates, Flare, and Sunburst as the bundle. For newsroom and merchandising workflows that need a character or product to survive many edits, this is the capability to test first.

Output sizes and poster workflows shift downstream tools

With arbitrary resolutions up to 3840x2160 and a max quality tier now in the menu, exports cover posters and merch use cases without a separate upscale step. Creators who previously bounced output through an Image Compressor or relied on Image Flipper for orientation fixes should expect the API to carry more of that work. For files that still need trimming or re-orientation before delivery, browser-side tools remain the simplest sidecar.

Practical moves for the week ahead

Practitioners should pick one product image and run it through both Flare and Sunburst, measuring latency and edit retention across three or more turns, then capture a 3840x2160 reference set with up to 16 inputs to confirm mask behavior. With a MIME Type Lookup on hand, exporting from the max tier into the right container stays predictable; if animated assets are still part of the pipeline, an Optimize Animated GIFs for Faster Web Pages pass keeps legacy channels healthy alongside the new API output.

Decisions on which model to standardize on can wait until both have been benchmarked against the team's actual reference sets.