OpenAI shipped ChatGPT Images 2.5 on September 8, alongside two new models in the API. The company is positioning this release as its best image model yet, with improvements concentrated in three areas: better instruction-following across multi-turn edits, less drift in the people and objects from reference photos, and faster generation. Latency is down as much as 50% compared with Images 2.0.
The rollout covers every tier of ChatGPT, ChatGPT Work, and Codex, pushed simultaneously to desktop, mobile, and web.
Sunburst and Flare
On the API side, the two model IDs are gpt-image-2.5-sunburst and gpt-image-2.5-flare. OpenAI's guidance on picking between them comes down to one line: use Sunburst for workflows where editing precision matters, and Flare for everyday fast generation. Early customers testing the models put Flare at two to four times faster than GPT-Image-2.
That split is unusual for an image model. Previous generations typically shipped as a single model, with speed and quality traded off through resolution and sampling steps. Fixing it into two separate IDs amounts to an admission that these two kinds of demand have grown large enough that sharing one set of weights no longer makes sense.
Doing the math
Pricing runs by token, not by image: image input costs $8 per million tokens, or $2 with a cache hit; image output is $30 per million tokens; text input is $5 per million tokens, or $1.25 cached.
For anyone calling the API, the practical effect is that cost floats with resolution and image complexity rather than following a flat per-image rate. A rough estimate: a 1024×1024 output image works out to somewhere in the low thousands of tokens, which against $30 per million puts the per-image cost at a few cents — not expensive on that side. What pushes the bill up is editing turns: multi-turn editing means feeding the previous round's output back in as image input, and that $8-per-million input cost starts stacking with every round. The $2 cached rate is aimed squarely at that scenario — repeatedly revising the same base image should hit the cache from the second round onward.
So the pricing treats the two use cases differently. Bulk generation that delivers on the first pass keeps costs predictable; workflows that treat the model as an image editor, revising one image a dozen or more times, rack up most of their bill on the input side, largely independent of how many finished images come out.
Draw a reference stroke
On the product side, several additions: Sketch lets users draw directly inside ChatGPT — invoked by typing @Sketch — with the drawing used as a reference; Templates gives common formats like posters, merchandise, and product shots a starting point; generated images can be annotated directly for targeted edits; and prompts can be shared so others can swap in their own photos and rerun them.
These features all point in the same direction: turning "describe, generate, start over" into "generate, then point at what to change." Both sketching and annotation hand positional information to the model, and position is exactly what's hardest to specify in a text prompt. Asking for an element to shift two centimeters left, or for the light in one spot to dim slightly, takes several sentences to describe in words — a stroke or a tap says it in one move.
Prompt sharing is a different kind of feature. It turns a prompt from a personal asset into a reusable template that others can pick up and rerun with their own material. This mechanism already has counterparts across the image community — prompt sites and model communities have been doing versions of it for a while — the difference here is that the platform now builds it in natively, without routing through a third-party site.
A number that doesn't add up
OpenAI's announcement cites a figure of 3 billion images, covering both ChatGPT Images and the GPT-Image family in the API. Different accounts of the number don't agree on what it measures — one version describes it as cumulative, another as weekly, two orders of magnitude apart. OpenAI's official page can't currently be checked word for word, so for now the number can only be read as "the scale is large" rather than used for any calculation.
The previous generation, Images 2.0, solved text rendering — its signature case was that text on menus and posters finally looked right. This generation targets editing, with its signature capability being that a subject still looks like itself after a dozen revisions. What the two generations have in common is that the real contest isn't over whether an image can be produced at all, but whether it can still be changed after it's done.
Sources: OpenAI announcement, CocoLoop, Simon Willison's blog, Unite.AI, 9to5Mac; model IDs, pricing, and latency figures follow the official announcement and its coverage, and per-image cost is a rough estimate based on published rates.