ChatGPT Images 2.5 arrived on September 8, 2026. OpenAI's launch announcement describes improved reference fidelity and editing consistency, with generation latency reduced by up to 50% versus Images 2.0. It also introduces two API models: Flare for everyday generation and Sunburst for more demanding edits, with longer generation times.
For a team building an image agent, the decision is where to spend the additional waiting and review time. A campaign visual that needs repeated corrections may justify a different model from a quick concept that will never leave an internal presentation. Start with the smallest workflow that produces something your team actually approves, then compare the alternatives on that job.
This is a source-based release analysis and a proposed evaluation plan. We have not run an independent image benchmark. The release is dated September 8; the API documentation and rates below were checked on September 9, 2026.
- 01Measure approved output.Include the time spent correcting, reviewing and exporting an asset when evaluating the upgrade.
- 02Test model choice by task.Use the same brief and source images to compare alternatives. Keep easy concepts separate from edits with strict preservation requirements.
- 03Budget beyond a single request.Retries, reference images and human review belong in the cost of delivering the finished creative.
- 04Keep approval attached to a file.Save the exact reviewed version and its brief so a later automated edit cannot silently replace the accepted asset.
01 — The releaseWhat changes for a creative team
The announcement adds Sketch, which turns a drawing into visual guidance; templates for common formats; comments placed on an image; and optional prompt sharing. OpenAI says the rollout covers all tiers of ChatGPT, ChatGPT Work and Codex across web, desktop and mobile. Those are announced access details, not a promise of unlimited use.
Our reading is that the value sits in making the brief less ambiguous. A rough layout can explain where a product belongs. A focused comment can distinguish changing one object from rebuilding a scene. For an operator commissioning creative, that may reduce the translation work between an intention and an instruction. Whether it reduces revisions is something to measure with the team's own material.
Keep a record of what must survive an edit: the approved product shape, packaging copy, subject position and reserved space for a headline. Otherwise an attractive new result can pass a casual review while quietly losing a business requirement. Better-looking output and a correctly executed brief are separate judgments.
02 — Model selectionStart with the cost of a failed edit
Flare and Sunburst should earn their place in a workflow through different tests. Our proposed starting point is to compare them where a mistake creates meaningful rework. The table below is an editorial routing plan, not a ranking from measured results.
| Work to test | Starting route | What would change the choice |
|---|---|---|
| Internal concepts and layout exploration | Flare first | Switch if extra retries erase the time saved. |
| A final product image with protected details | Compare both on the same edit | Prefer the route with less product drift and less reviewer intervention. |
| A sequence of revisions to approved creative | Include Sunburst in the pilot | Check whether earlier accepted changes survive later requests. |
| A large set of channel variations | Trial a small representative subset | Expand only after export, crop and review requirements pass. |
Avoid selecting the model solely from its strongest demonstration. Your business may need an ordinary bottle photographed correctly more often than an elaborate illustration. Include mundane cases: a pale label on a pale background, a small disclaimer, a narrow crop and a change requested after three earlier approvals.
Ask reviewers to score outputs without seeing the model name. Record the defect they found, not just whether they liked the image. That creates a useful routing rule: a particular job fails because packaging text changes, rather than because one model feels less capable. Our guide to testing a model on your own traffic explains how to make that comparison relevant to a production decision.
03 — The billBoth models share token rates, not necessarily a bill
OpenAI's API pricing page lists identical standard rates for the two models. Prices below are US dollars per one million tokens. Text input is the written prompt; image input is supplied visual material; image output is the generated visual content.
| Token category | Flare | Sunburst |
|---|---|---|
| Text input | $5.00 | $5.00 |
| Cached text input | $1.25 | $1.25 |
| Image input | $8.00 | $8.00 |
| Cached image input | $2.00 | $2.00 |
| Image output | $30.00 | $30.00 |
The image generation guide warns that the models can use different token counts at the same quality setting. It also notes that Responses API requests add the coordinating model's token usage to image-generation costs. Use explicit size and quality settings when estimating; automatic settings make a fixed estimate less meaningful.
A straightforward costing rule is to multiply each category's measured token count by its rate and divide by one million. Keep cached and uncached input separate. Then sum every attempt required for the job. One thousand image output tokens would cost $0.03 at the listed rate; that is a unit calculation, not an estimate that an actual image uses one thousand tokens.
For a business decision, add reviewer time and divide total expenditure by accepted assets. An unsuccessful attempt still consumes resources. So does the designer who must repair a distorted label. A lower generation bill can lose its advantage when the review queue grows; our analysis of human review costs in model selection develops that comparison.
Track cost per approved asset and elapsed time to approval. Keep generation latency as a diagnostic underneath them. A faster request helps only when it moves the finished job forward.
04 — ImplementationKeep image generation inside a bounded job
The official Flare model card identifies gpt-image-2.5-flare and the dated snapshot gpt-image-2.5-flare-2026-09-08. It accepts text and images and produces images. Supported quality choices include low, medium, high, xhigh, max and auto. Pinning a snapshot gives a pilot a recorded model version.
The Images API supports direct generation; the Responses API can delegate image work through its image generation tool. The example below uses the direct route and saves the returned base64 image. It illustrates the request boundary and has not been executed for this article.
import base64
from pathlib import Path
from openai import OpenAI
client = OpenAI()
result = client.images.generate(
model="gpt-image-2.5-flare-2026-09-08",
prompt=("Create a square editorial illustration of a ceramic mug "
"on a desk, with open space above it. Do not add text."),
size="1024x1024",
quality="medium",
)
if not result.data or not result.data[0].b64_json:
raise RuntimeError("No image returned; keep the job unapproved")
Path("mug-concept.png").write_bytes(
base64.b64decode(result.data[0].b64_json)
)In an application, store a job identifier, the brief, source-file references, model version and resulting file together. Keep the API credential on the server. Put an explicit limit on attempts and total spend. If the system cannot produce an acceptable result within that allowance, return the remaining problem to the person responsible for the asset.
Give the agent a concrete editing instruction and a list of protected details. After each change, compare the result against the last approved version. Save both versions so a reviewer can inspect the difference. An agent that repeatedly regenerates from its own newest image can gradually move away from the original brief without producing an obvious error message.
05 — The first pilotTest the revision loop your team already uses
Choose one recurring job with a named approver. For example, use a product campaign that needs a background change followed by space for a headline. Work with source images your team is authorized to use. Write the acceptance criteria before generation: which objects can move, which copy must remain exact, and what file the recipient needs.
Run the same sequence against each candidate. First generate the initial composition. Next request one constrained change. Finally make a different change that should preserve the first. Review the sequence against the original references as well as the preceding image. This catches a common evaluation gap: judging each output in isolation misses whether the workflow remembers prior requirements.
Record generation time, attempts, correction requests, review minutes and the final disposition. Separate rejected outputs from accepted ones instead of dropping failures from the report. Repeat representative cases because a single attractive result provides little evidence about repeatability. Publish the sample size and settings internally with the result, so colleagues can understand its limits.
Set the decision rule before looking at results. A team might require the new route to preserve every protected detail and reduce review effort on the chosen job. Another may accept a longer wait if the final export needs no manual repair. Those are business preferences to establish with the approver; the release announcement cannot choose them for you.
Expand only to nearby jobs that share the same constraints. Passing an illustration task does not establish readiness for packaging, and passing a square composition does not establish that every channel crop works. Keep the existing production route available while the new workflow earns broader responsibility.
06 — DeliveryPreserve the origin record and review the actual asset
OpenAI's Images 2.5 system card describes C2PA metadata and Google DeepMind's invisible SynthID watermarking across ChatGPT, Codex and the API. C2PA carries provenance information; the watermark adds another signal of origin. OpenAI describes a layered approach rather than a single complete solution.
A production team should retain the original generated file and record subsequent edits. Check what survives the actual export and delivery path instead of assuming that a downloaded, resized and recompressed asset carries the same information. Provenance helps explain how a file was created; it does not establish that the depicted product has the pictured features or that every statement on a label is correct.
Bind approval to the exact deliverable, destination and intended use. If an agent changes the copy, subject or composition after approval, send that version back through review. Our guide to reviewing an agent's changes before publication applies the same principle to the final handoff. This is especially useful when image generation feeds directly into a publishing workflow.
07 — The decisionMake the upgrade earn its place in the workflow
Buy a better path to an approved asset.
Start with one job, a clear brief and an approver. Compare model routes on preservation, revisions and total cost, then expand the route that meets the actual delivery requirements.
The practical opportunity is fewer interruptions between an idea and its finished creative. The evidence that matters will come from the work your team can repeatedly accept.