ChatGPT Images 2.5: what creators and founders should build

Yes, ChatGPT Images 2.5 has been released. OpenAI announced it on September 8, 2026, with two API variants: GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. OpenAI describes improved reference fidelity, focused edits and consistency across revisions. Flare is positioned for general use; Sunburst targets more precise creative work with longer generation times. These are vendor claims, not results from our arena. OpenAI announcement.

The useful question for a founder is which part of a real production job becomes easier. A strong first image is helpful. A product image that survives three requested changes and still represents the actual product is much more valuable.

My recommendation is to test one complete commission before reorganizing your stack. Keep the approved reference, make the intended edits, export the required formats, and count the repair work. The reference-to-campaign guide includes the prompts and acceptance checks.

Which model should you try?

Do not declare a winner from separate prompts. Use the same reference files, edit instructions and target dimensions. Give each workflow the same maximum number of attempts. Include an ordinary production baseline: your current model, a template, or a designer's usual process.

This release has no first-party image results in the Model Arena yet. The published protocol is a test specification, with status: not_run and lastMeasured: null. It is suitable for preparing a comparison, not citing a success rate.

What becomes possible for a creator?

Start with one approved visual and make it a reusable source for the project. A course creator could develop a scene that supports a lesson cover, chapter illustrations and social excerpts. A musician could build a consistent visual world across a cover, release announcement and stills for a later video workflow. An author could establish a character reference and evaluate continuity across several story moments.

These are proposed uses. The value depends on whether the actual output keeps the identity and composition the creator approved. A generated still does not establish that animation, sound or an editable design project has been produced.

The workflow should keep four records: the source reference, the approved version, the requested change and the exported asset. That small amount of structure prevents a common problem: every revision becomes a fresh interpretation of the brand.

What becomes possible for a founder?

A product launch is a good test because it contains several different constraints. The product must remain accurate. The message must remain legible. Each placement has a different crop. A last-minute offer change should preserve the rest of the design.

Commission a small launch package: a landing-page image, square social asset and portrait announcement. First approve a direction. Then request one change at a time. Keep exact prices, dates and legal copy in editable text layers when correctness matters. Inspect every exported placement on its intended device.

Measure time to the approved package, including manual repair. A workflow that produces attractive concepts quickly but requires repeated identity corrections may still be a poor fit for your business.

How this should work across our products

The following are product directions to build and validate. Links lead to each brand's existing home; they do not promise that an Images 2.5 feature is already available there.

GenCreator should own the repeatable campaign job. Arcanea should own continuity within a creative world. Starlight should retain the project records and help choose an appropriate execution path. FrankX should explain the workflow and publish inspectable evidence. That gives each destination a reason to exist.

The strongest next feature is an approved asset library with revision history. It can answer: which source produced this image, which version did we approve, and what must remain unchanged? A prompt collection alone cannot answer those questions.

Build the evidence into the workflow

Evaluate six jobs: a campaign scene, selective background edit, exact-copy change, three-step revision sequence, character continuity and a portrait derivative. Our guide explains the scoring anchors. The JSON protocol lets an agent read the same instructions without scraping a page.

For each attempt, retain the exact provider model ID, input checksums, request settings, elapsed time, usage where returned, output checksum and reviewer decision. Keep failed attempts in the denominator. Do not publish private reference images without permission; a private run can retain hashes and publish an appropriately redacted description.

Use a blind human review for visual acceptance. Identity and mandatory copy are hard requirements. A beautiful image with the wrong product details fails the job. If there are too few observations to compare reliably, publish the examples and limitations instead of a leaderboard.

What will it cost?

At the September 10 check, both variants list image-token rates of $8 input, $2 cached input and $30 output per million tokens. Text input is $5, with cached text input at $1.25 per million tokens. These are token rates, not a fixed price per image. OpenAI pricing.

Total cost depends on the actual request, output and repeated attempts. Track API cost and human correction time separately. Then compare cost per approved asset, including the spend on rejected attempts. The model registry JSON keeps image pricing separate from text-model pricing so agents can preserve that distinction.

Make the next step useful

If you want to create something now, use the reference-to-campaign guide. If you are selecting a model, inspect the LLM Hub. If you want evidence, use the arena receipts, where a missing measurement remains a missing measurement.

For agents and product builders, the workflow routing manifest connects intent to those resources and marks proposed product features explicitly. It can support a handoff without sending every visitor to the same sales destination.

The practical first move is small enough to finish: take one real brief through reference selection, generation, revision, review and export. Retain what worked. The reusable decisions and honest results from that process are what make the next project better.