OpenAI discounts background work that can wait
OpenAI has introduced discounted completion windows for Responses API background work, offering a lower model-token price when an application can wait for its initial response. Its 18 November…
OpenAI has introduced discounted completion windows for Responses API background work, offering a lower model-token price when an application can wait for its initial response. Its 18 November announcement highlights GPT-6.2 Astra and Sol and describes the feature as a public beta.
The current documentation gives a two-hour window a 30% discount below Standard model-token rates and an eight-hour window a 40% discount. The default, soonest, carries no discount. The guide also lists GPT-6.1 Terra and GPT-6 Luna among supported models.
The clock applies to the first response
The selected window starts when the API accepts the initial request and targets that response. It is not a promise that an entire multi-step agent run, including external tools and follow-up requests, will finish within two or eight hours.
Eligible follow-up requests can receive the token discount too. The documentation says to reuse the window, continuation ID, model and project to qualify for higher queue priority. This still requires the application to run tools, handle responses and manage failures.
Where a merchant app could use it
An overnight document review or a support-ticket analysis needed for the next morning is a plausible fit. Checkout assistance or a customer waiting for an answer is not. These are suggested applications, not packaged Shopify features supplied by OpenAI.
Set background mode, choose the completion window and use Standard processing as documented. Save the response ID, receive and verify completion webhooks, and retrieve the response. If webhooks are unavailable, poll that same response ID instead of resubmitting the job.
For independent requests that can all be submitted up front, the guide points to the Batch API. Budget for tool calls, storage and application work separately: the advertised percentage applies to eligible model tokens, not every cost in the workflow.
Also review data handling before sending sensitive material. The guide states that background mode temporarily stores response data even with store set to false or Zero Data Retention, and says temporarily stored request content is encrypted at rest.
More from AI
See all →
