共有:
Image Generation Update

OpenAI Ships GPT Image 2.5
With Two Speed Tiers: Flare and Sunburst

Speed and editing precision have long been a trade-off in image generation. OpenAI's new model generation tries to break that assumption entirely — by offering two purpose-built models instead of one compromise.

AI Navigate Editorial2026.09.156 min read
FLARE SUNBURST prompt single pass up to 50% faster edit 1 edit 2 subject & background preserved
01
Why Now

Why now — the limits of
latency and "whole-image redraws"

Teams that put image generation into production have lived with two persistent frustrations. The first is latency: for use cases like social content or ad production, where large volumes of images need to go out fast, a few seconds per image can define the whole workflow's feel. The second is imprecise editing — asking to change just a logo's color or swap out part of a product package, only to have the model redraw the entire background and layout along with it. As TestingCatalog reports, competitors including Midjourney have been racing to ship editing-focused models this same month, making both problems an industry-wide focal point.

On September 8, 2026, OpenAI posted "Introducing ChatGPT Images 2.5" on its community forum, adding two new API model routes: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Per the official announcement thread, Flare becomes the new default model, while Sunburst is offered in parallel as a precision-first option. Rather than forcing a single compromise, the release lets teams choose the model per use case.

Flare (fast, new default)Sunburst (precision editing)
Up to 50% lower latency vs. GPT Image 2Longer generation time, prioritizes precision
Matches GPT Image 2's image qualityTargets only the requested element
Built for social content, visual search, rapid prototypingKeeps composition and brand treatment stable across repeated edits
Optimized for high-volume workloadsOptimized for premium, brand-focused production

Speed or precision — pick one, no more.
This release tries to erase that choice altogether.


02
Under the Hood

How Sunburst edits
with surgical precision

It answers the "whole-image redraw" problem that has broken backgrounds and logos on every edit pass — by changing only the targeted region.

BEFORE: full redraw Background and logo shift every time Sunburst AFTER: targeted edit only Subject, background, brand treatment kept
FIG. Typical editors redraw the whole image on every instruction; Sunburst rewrites only the targeted region, holding subject, background, and layout stable across the edit chain.
Up to 50%
Flare's latency cut vs. GPT Image 2
2 routes
New API models (flare / sunburst)
All tiers
Sketch tool rollout

This design pays off most in multi-step edit chains — changing only a logo's color, or swapping one part of a product package, across several consecutive passes. Sunburst targets only the requested element and keeps the subject, background, layout, and brand treatment stable across the whole edit chain. TestingCatalog's technical writeup notes that this precision editing is aimed squarely at premium, brand-focused visual workflows.

The same release adds a new "Sketch" tool to ChatGPT Images, which turns a rough hand-drawn sketch into a finished AI image, rolled out to every ChatGPT tier. Starting from a rough sketch removes the friction of putting an idea into words as a prompt in the first place.

03
Who Benefits

Who benefits, and how

The value lands differently depending on role. The two models aren't competing options — they're a division of labor.

Engineers & product teams

For high-volume calls — visual search, rapid prototyping — Flare's lower latency feeds directly into both API cost and UX. Since the default is switching to Flare, code that doesn't pin a model ID explicitly will change behavior, so check that first.

Designers & creative teams

The frustration of touching up a small detail without breaking a logo or product shot is eased by Sunburst. The new Sketch tool also opens a sketch-first workflow, letting ideas move straight into an image without first being translated into a prompt.

Business & marketing leads

Choose Sunburst for production that must hold brand guidelines across repeated iterations, and Flare for high-volume social output — which means a new cost-allocation call has to be made per workload.


04
What's Next

What to do next

  1. If your API calls don't pin a model ID, check the impact of the default switching to Flare before it reaches production.
  2. For production with long edit chains — banners with logos, package swaps — trial Sunburst and compare quality against your current pipeline.
  3. Simulate cost per workload (high-volume vs. precision editing) before committing to a full switch.
05
Risk & Limits

Not unqualified good news

This isn't a free upgrade. Multiple outlets report that API pricing for the new routes has roughly doubled compared to GPT Image 2 — the pricing of record should be confirmed against OpenAI's own announcement thread. For high-volume workloads, that price increase could offset Flare's speed gains, so this isn't simply an upgrade with no downside.

It's also unclear from public information how well Sunburst's targeted editing holds up under more complex instructions, such as editing multiple objects in a single pass. With competitors including Midjourney shipping editing-focused models around the same time, how the precision claims hold up will need to be judged by real-world use.