OpenAI releases GPT-Image 2.5 Flare and Sunburst API models

• Flare is the default API option for most workloads with up to 50% lower latency than GPT Image 2 under OpenAI's model-selection guidance, while Sunburst trades longer generations for tighter edit control.
• ChatGPT Images 2.5 reached ChatGPT, ChatGPT Work, and Codex across desktop, mobile, and web, as OpenAI's rollout notice announced.
• Arena placed Sunburst first and Flare second across text-to-image, single-image edit, and multi-image edit leaderboards, with Arena's announcement reporting the largest gain in multi-image editing.
• The release centers on retaining the parts of an image that an edit did not target, which OpenAI's editing demo illustrates across product, background, and copy changes.

OpenAI says more than 3 billion images are created weekly across its ChatGPT and API products. Its image-prompting guide carries a decidedly unmagical caveat: repeated edits can still alter details meant to remain fixed, and pixel-identical regions may need compositing. The system card says outputs carry C2PA metadata and invisible watermarking.

• ChatGPT Images 2.5 adds Sketch, templates, in-image comments, and prompt sharing, according to OpenAI's product announcement.
• The API adds gpt-image-2.5-flare and gpt-image-2.5-sunburst; OpenAI's API guidance calls Flare the default for most applications.
• Both variants generate and edit images, including transparent-background workflows, in OpenAI's image prompting documentation.
• OpenAI's launch post contains no new API price table or rate-limit change.

First-party

• Generation latency: Images 2.0 baseline → up to 50% lower with Images 2.5, up to -50%, per OpenAI's launch post.

Third-party evaluators

• Text-to-Image Arena score: GPT Image 2 (medium) 1,381 → Sunburst 1,421, +40 points, per Arena's announcement.
• Single-Image Edit Arena score: GPT Image 2 (medium) 1,461 → Sunburst 1,520, +59 points, per Arena's announcement.
• Multi-Image Edit Arena score: GPT Image 2 (medium) 1,454 → Sunburst 1,535, +81 points, per Arena's announcement.

Customer-reported

• Generation speed: GPT Image 2 → Flare at two to four times the speed, +100% to +300% throughput, in ManusAI's evaluation.

Sunburst carries the deliberate tradeoff: OpenAI's API guidance describes it as the precision option with longer generation times.

OpenAI's image-prompting guide says repeated edits can still move details intended for preservation. It tells developers to evaluate an entire edit sequence, not merely individual edits, and to composite an approved edit into the source when a region must remain pixel-identical.

A nonstandard animation test by chetaslua put center jitter at roughly 5 to 8 pixels in GPT Image 2 and 0.8 pixels in 2.5, but still measured about 3% size jitter in 2.5.

A side-by-side stability comparison

Failures remain visible in early use. In one result, nicdunz found unwanted text scattered through a generated scene, and another nicdunz post reported that every generated image in a batch had a visible problem.

The public distinction is model class and optimization, not a disclosed new architecture.

• Flare: OpenAI's guide calls it the small model, tuned for speed, with quality comparable to GPT Image 2.
• Sunburst: The same guide calls it the base model, tuned for quality above GPT Image 2.
• Shared capabilities: Both accept reference images, support generation and edits, preserve subjects more reliably, and handle transparent backgrounds.
• Undisclosed details: Neither the launch post nor the guide specifies architecture, context window, training recipe, or a price differential between the two models.

• A reference-image CLI example on Simon Willison's Weblog passes a source image with -i to Sunburst, then asks the model to add a raccoon scientist to an existing graphic.
• Higgsfield's lamp-moving test asks for one physical change, moving a lamp to the right, and shows the reflected lighting changing with it.
• A tightly specified, eight-second pixel-art sequence in Higgsfield's comparison shows 2.5 retaining the supplied character design more closely than GPT Images 2.0.
• sound4movement generated a stick figure plus nine successive cartwheel frames, then had the model assemble them into a flipbook GIF.
• One early workflow report from aibuilderclub says ChatGPT selects the model itself; Flare and Sunburst are selectable at the API layer.

GPT Images 2.0 and 2.5 on a constrained pixel-art animation

Day-one distribution extended beyond OpenAI's own clients.

• Vercel's AI Gateway changelog exposes openai/gpt-image-2.5-flare and openai/gpt-image-2.5-sunburst, matching Vercel's AI Gateway announcement.
• OpenRouter listed both models, and its Images API documentation covers reference-image inputs, output controls, and streaming where a provider supports it.
• fal announced both targeted editing and long-session consistency claims for its hosted release.
• Replicate listed GPT Image 2.5 for precise editing and multi-turn consistency.
• Pika added both models to API Club, including transparent-background output.
• Teknium said GPT-Image-2.5 had reached Hermes Agent through Codex subscriptions and fal.