Discuss your project
Artificial Intelligence

ChatGPT Images 2.5: measuring the cost of an approved visual, not generated volume

ChatGPT Images 2.5 emphasizes editing and iterative work. Measure the cost of a validated visual according to the brief, rework, approval and provenance.

Visuel validé selon une grille de qualité et de provenance.

ChatGPT Images 2.5 emphasizes editing and iterative work. For a marketing team, however, the useful measure remains unchanged: what does a visual that is genuinely approved, matches the brief and can be used in a campaign cost?

What is announced for creative work

OpenAI presented ChatGPT Images 2.5 on 8 September 2026. The announcement cites localized comments for requesting a precise edit, Sketch, which uses a drawing as a reference, and the Flare and Sunburst models accessible through an API (application programming interface). These features and claims of accuracy are publisher statements: they prove neither a given account's access, brand fidelity, nor a production gain for a team. [OpenAI announcement]

This article's illustration is an original editorial creation. It is neither a comparative test nor an output attributed to ChatGPT Images 2.5; the version of the generation tool is not disclosed.

Writing the brief as controllable criteria

A creative requirements document, or brief, describes the invariants: shape, material, framing, dimensions, text area, permitted and prohibited elements. For a pilot, use a fictional, unbranded geometric object. Give each criterion an owner and validation evidence. Reuse rights require a specific review; they do not automatically follow from a tool.

CriterionExpectedOwnerEvidenceDecision
Shape and compositionTo be completedTo be assignedNot measuredTo be approved
Text and readabilityTo be completedTo be assignedNot measuredTo be approved
Origin and rightsTo be completedTo be assignedNot verifiedTo be approved
Visual validation grid: fictional product, brand, readability and provenance.
The brief becomes a set of debatable and verifiable criteria.

Calculating the cost of an accepted visual

Count preparation, generation, selection, correction, review and integration. Set the conventions: active human time, billed tools and validation costs are included; media buying, distribution and commercial effects are excluded if the pilot does not measure them. The proposed formula is: documented production and validation costs / number of accepted visuals. When no visual is accepted, the result is undefined: it must neither be replaced with zero nor used to promise savings.

StepTime or costIncluded?Evidence
PreparationNot measuredTo be decidedTo be attached
Generation and selectionNot measuredTo be decidedTo be attached
Rework and reviewNot measuredTo be decidedTo be attached
IntegrationNot measuredTo be decidedTo be attached
Steps to include in the cost of a genuinely accepted visual.
Generated volume does not replace the number of approved deliverables.

Testing use without inventing a commercial effect

A compliant image is not proof of better conversion. A campaign test must define exposure, objective, variants and a stopping rule before it is carried out. Until it has taken place, no marketing result can be attributed to the creative flow.

Keep the brief, the origin of the elements, versions and the approval decision. To complete this record, see our article on AI content provenance. The pilot a team may decide on remains small, non-sensitive and documented: it measures rework and accepted visuals, while leaving product access, rights and marketing effect open questions.

Share this article