Heimstaden’s cost of producing a virtually staged property image in the Netherlands fell from approximately €50 to less than €0.10, according to a customer story OpenAI published on November 3, 2026. For letting teams, the next question is the cost of an image ready to publish.
Virtual staging furnishes rooms digitally, not physically. OpenAI’s account estimates €130,000 in annual savings potential for Heimstaden in the Netherlands—not a reported saving already achieved.
The case study says employees review AI content and approve decisions. It does not itemize review, training, implementation or software costs, or identify the image model and measurement period. It is a supplier-published account, not an independent audit; equivalent savings elsewhere and faster lettings remain unestablished.
The reported image cost is not a price quote
OpenAI lists ChatGPT Enterprise at custom prices, with both credit-based and token-based agreements available. The customer’s reported image cost is therefore not a generally purchasable rate.
Under eligible token-based Enterprise agreements, usage charges sit alongside any contracted seat fees. The Enterprise image rate card also explains that request costs depend on text input, reference images used for editing, and image output. Dimensions and quality affect consumption.
For a budget, use your actual agreement and bills. Include the staging workflow’s share of software costs, but do not assign an entire company subscription to it or count usage twice when the plan already includes it.
Decide what counts as a usable image
For a property-marketing pilot, measure the cost per approved image, including the work needed to make it ready for its intended listing.
A convincing picture is not enough. OpenAI’s image-generation documentation acknowledges that models can struggle to position elements precisely in layout-sensitive scenes. That is a general limitation, not a finding about Heimstaden’s images. A proposed review should compare the edited image with the original photograph, including the room’s windows, doors and fixed features.
Where the image will appear also matters. The Dutch property platform Funda’s publication rules prohibit AI-generated images as the main listing photo, except for new-build projects. Additional images that materially change reality must disclose the alteration within the image and represent the property realistically.
Those are Funda-specific rules, not a universal rule or evidence that Heimstaden lists there. The budgeting implication is that an image suitable as a labelled illustration may not replace the main photograph. Set the destination, photo role, accuracy and disclosure requirements before counting output as approved.
Count the whole production process
A useful comparison is:
Cost per approved image = all attributable production costs ÷ distinct final images approved for their intended use.
Count charges for rejected attempts as well as accepted ones, plus staff time for preparation, review, revisions and delivery. Add allocated software costs and a stated share of setup and training. Compare the same number and type of required images, at the same quality and delivery standard, with the existing service.
Include necessary source photography on both sides. If it is an unchanged shared cost, show it separately and explain why it cancels out of the savings comparison. Keep tax and currency assumptions consistent.
Illustrative calculation—not Heimstaden’s results or a supplier quote. Suppose a batch yields 100 approved final images from 300 generation attempts. Every input below is assumed; the €0.10 per attempt is not the customer’s reported less-than-€0.10 figure or an advertised tariff.
Assumed cost for the batch | Amount |
|---|---|
300 attempts at €0.10 each | €30 |
1,000 staff minutes across the full process at €30 an hour, including employment costs | €500 |
Allocated software costs, separate from attempt charges | €100 |
Allocated setup and training costs | €50 |
Total for 100 approved images | €680 |
The arithmetic is €30 + €500 + €100 + €50 = €680, or €6.80 per approved image. An extra five minutes of work per approved image at the assumed €30 an hour adds €2.50, bringing the example to €9.30.
This simplified example excludes unchanged photography, taxes and any other unlisted costs. Add them where applicable. The setup allocation spreads a cost across output; it is not necessarily the upfront cash payment. State the period and expected volume used for that allocation.
The example shows why staff time matters, not how much a particular business will save. It should not be subtracted from Heimstaden’s approximate €50 baseline: the scope of that service is not sufficiently specified for a like-for-like comparison.
Check the cost of a complete listing, too
A pilot can record each source photo, the attempts it produces, review minutes, charges and the final acceptance decision in a spreadsheet. Count distinct required deliverables, not every variant. If no images pass review, there is no usable unit-cost result.
Track cost per completed listing alongside cost per approved image. Producing more variants could reduce the average cost per image while increasing total spending. Keep any extra output separate from the comparison with the original job.
Define success before examining the result: lower spending for the same deliverables, less staff time, or additional useful output. RohitAI’s analysis of human review and the choice of metrics explores why checking the arithmetic is different from checking what a calculation measures.
Finally, separate avoided invoices from time freed for other work. Released staff capacity can be valuable without reducing payroll. Expand production when comparable, approved listings show a worthwhile benefit after review and setup costs—not simply because generating another image is cheap.