Price the accepted image, with its model and retries attached
Retired models and changing settings make old per-image quotes unreliable. Track the current request configuration, input charges and attempts per accepted asset.

A four-cent image price is meaningful only when it identifies the model and the request that produced it. A square draft can have a different cost from an edited banner. The cost of an accepted asset also depends on the attempts that preceded it.
For an image-generation application, the useful unit is the cost of producing an asset that meets the user's task. Check current model availability first. Record the request configuration with the attempts needed before acceptance.
Check the model before comparing a remembered price
OpenAI's deprecation schedule records the removal of DALL-E 3 from the API on 12 May 2026. It also records the shutdown of gpt-image-1 on 23 October 2026. Older preset advice built around those API models needs an explicit migration.
As checked on 14 November 2026, the same schedule gives gpt-image-1.5 a shutdown date of 1 December 2026. A model that still answers requests today can therefore require a near-term replacement. Keep that date in the maintenance plan for an application using it.
OpenAI's current documentation describes the GPT Image 2.6 Sunburst and Flare models. Selecting a replacement requires testing the actual task with that model. A familiar quality label can survive a model change while the resulting images and token counts change.
Record the complete model identifier in the generation history. A preset called Pro leaves the underlying implementation invisible unless its settings are stored elsewhere. The operator needs to reconstruct which request produced a result after the default changes.
Estimate the request using the actual settings
The current image-generation guide provides a calculator for a selected model. Its estimate also depends on the chosen quality setting and size. It also explains that the estimate concerns image output cost, with text and image inputs accounted for separately.
For GPT Image 2.6, the documented quality choices include settings beyond high. The guide also describes complexity as a separate control over planning before rendering. A product should preserve the meaning of each setting when it builds its own labels around them.
OpenAI's standard pricing table, checked on 14 November 2026, gives GPT Image 2.6 Sunburst image output at $30 per million tokens. Image input is $8 per million; text input is $5 per million. Those are token rates, with different treatment for cached input and eligible processing modes.
For illustrative arithmetic, 2,000 output tokens at that output rate cost $0.06. The request's input charges still need to be added. The assumed 2,000 tokens are an example for the calculation, with no claim that every image uses that amount.
An edit can include reference images that contribute input usage. Preserve that usage beside the output count when reconciling the request. Comparing only the visible dimensions would leave part of the bill outside the explanation.
Measure the cost per accepted asset
A low-cost attempt can be useful for exploring a composition. Its value depends on whether the draft helps the user reach an acceptable result. Repeated attempts that change the wrong element can consume both generation spend and review time.
Use a fixed task when comparing presets. For a banner, define the required composition and the text that must remain legible at its display size. Keep the evaluation criteria available to the person reviewing the outputs.
In a hypothetical comparison, a five-cent request repeated three times costs fifteen cents. A twelve-cent request accepted on the first attempt costs twelve cents. Those assumed prices demonstrate the arithmetic of retries. They provide no performance ranking for a named model.
Record unsuccessful attempts as part of the task. If the application deletes them from the history, the final accepted file can appear cheaper than the process that produced it. A production cost calculation needs the generation activity that contributed to that outcome.
Review time can also affect the choice. A slightly more expensive request may reduce inspection work for one task while providing no benefit for another. Measure the difference with the kind of asset the user intends to publish.
Put an honest estimate beside the action
A generate button can show an estimated request cost when the application knows the selected configuration. State what the estimate covers. If it excludes reference-image inputs or later retries, keep that limitation beside the displayed amount.
Store the settings with the estimate accepted by the user. After completion, compare the returned usage with that estimate. A discrepancy then has a specific request to investigate instead of an unexplained difference in the monthly invoice.
Keep automatic retries visible in the product's cost model. A retry policy that makes the interface appear smoother can still increase usage. Establish a bounded retry rule and retain the result of each attempt so the operator can reconcile the total.
Subscription benefits belong in a separate account check
OpenAI's billing guidance separates ChatGPT billing from API billing. Current API-credit benefit documentation also describes credits that eligible Pro and Business subscribers can claim for a linked API organisation. Plus is outside that offer's eligible plans.
An eligible subscription needs activation of that benefit and confirmation of the receiving organisation. A product's API request still creates API usage, with the account's applicable credit balance determining how the charge is funded.
For the next preset change, retain a small set of representative tasks with their accepted outputs and total usage. Run the candidate configuration against those tasks before changing the default. The comparison will then concern the cost of the result the application actually promises to deliver.