Only 4.5 months in.
Image generation already jumps to a new generation
GPT Image 2 only shipped this April, and it was still the shiny new thing — until OpenAI announced its successor, ChatGPT Images 2.5, on September 8. Generation latency is cut by up to 50%, a new "Sketch" tool turns rough drawings into finished art, and two new API models arrive for developers. Image models are now iterating on the same cadence as text LLMs.
Barely half a year after launch,
the next model already arrived
On September 8, 2026, OpenAI announced its new image model, ChatGPT Images 2.5. According to the official announcement, the improvements come down to three things: more natural lighting and textures, editing that better preserves the subject from a reference photo, and more reliable instruction-following across multiple turns of editing. This isn't just a new visual style — it's an upgrade aimed at not breaking down as edits pile up, which matters a lot in day-to-day production work.
The predecessor, GPT Image 2, reached the API and Codex on April 21, 2026, and ChatGPT the next day as "ChatGPT Images 2.0." At the time, TechRadar's review rated it ahead of Google's "Nano Banana 2." That means a successor arrived just four and a half months later — a sign that image-model release cycles have caught up to the roughly six-month cadence of text LLMs like the GPT-5 series.
| GPT Image 2 / Images 2.0 (Apr 2026) | ChatGPT Images 2.5 (Sep 2026) |
|---|---|
| Editing centered on text instructions | New "Sketch" tool for drawing a reference |
| Background removal needed manual work | Remove BG and other tools built into the editor |
| Multi-turn edits could degrade over time | Improved to hold up across repeated edits |
| One API model family | Two models: flare and sunburst |
Image-model releases have already
settled into a six-month cycle.
Images 2.5
by the numbers
The official announcement disclosed few hard figures, but the ones that made it out matter to developers.
Two new model IDs were added for developers: gpt-image-2.5-flare, a standard model carrying the same quality and speed gains as Images 2.5, and gpt-image-2.5-sunburst, a high-precision variant that trades longer generation time for more detail. The OpenAI Developer Community announcement also spells out when to reach for each. The rollout covers ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web.
Designers, marketers, engineers —
what changes for each
Rather than a feature list, here's what actually changes by role.
Designers: sketch it, then finish it
A new "Sketch" tool lets you draw a rough reference directly in ChatGPT and type "@Sketch" to turn that line drawing into a finished image. The editing toolbar also gains Markup, Comment, Remove BG, Erase, and Resize — Remove BG exports a transparent PNG of just the foreground, cutting out a step of manual compositing.
Marketers: templates and consistency
New templates for common formats like flyers and product photos mean less prompt engineering from scratch. Edits now hold up better across multiple turns, so it's more workable for producing a series of ads that keep the same tone and look. Being able to share the prompt behind an image also makes it easier to hand work off internally.
Engineers: two API models to choose from
Two new model IDs, gpt-image-2.5-flare and gpt-image-2.5-sunburst, let you pick speed or detail. The official announcement cites up to 50% lower generation latency versus Images 2.0, which is worth revisiting your batch-generation cost and latency budget over.
Worth doing this week
Try it yourself
Test "@Sketch" in ChatGPT and see how well it turns a rough draft into a finished image for your use case.
API users: compare the two models
Run the same prompt through flare and sunburst to decide which one fits your use case — speed or detail.
Audit existing assets
Find images you could re-export as transparent PNGs with Remove BG, and list candidates to swap into ads or decks.
A few reasons not to celebrate too soon
The only hard numbers in the official announcement are the up-to-50% latency cut and the two new API models — the quality improvements are described in qualitative terms like "more natural textures" and "better preserves the subject," without benchmark scores to back them up yet. It's not yet clear from primary sources whether API pricing carries over from GPT Image 2 unchanged, or how sunburst's longer generation time will affect cost in practice.
Third-party comparisons against Google's "Nano Banana 2" haven't caught up yet, so it's reasonable to wait a few weeks before drawing conclusions about real-world performance. Don't assume the announcement means it's automatically the best choice for every use case — test it against your own workflow before switching over.