OpenAI discounts background mode by up to 40% for 8-hour jobs

OpenAI has added two completion-window tiers to background mode in the Responses API. Developers set background: true in a request, then pick 2h or 8h through background_options.completion_window. Model token rates drop 30% and 40% respectively compared with Standard. If the parameter is left out, requests default to soonest and are billed at the regular price. Third-party developer sites recorded the launch date as November 17, and OpenAI's developer documentation now spells out the full rules.

Background mode itself is not new. It turns long-running requests into asynchronous ones: the client receives a response ID, can disconnect, and later retrieves the result through a webhook or by polling. What is new is a “pay less if you are willing to wait” option. The documentation's examples are down to earth: the 2-hour tier suits analyzing support tickets that piled up over the morning before an afternoon triage meeting, while the 8-hour tier suits dropping a batch of documents in for overnight review and checking the results the next morning.

Illustration: three belts carry parcels, with coins on the two longer belts marking OpenAI background task discounts

What it takes to qualify for the discount

The discount is currently in public beta and covers only GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. The restrictions are spelled out in detail:

  • service_tier must be set to default, meaning Standard processing. If it is set to auto, the discount applies only when the API actually selects Standard;
  • 2h and 8h requests cannot use streaming, so stream must be omitted or set to false;
  • These two tiers share a rate-limit pool with the Batch API, and input tokens count against it at submission time;
  • The window is a target: after the API accepts a request, it tries to finish within the window, and automatically retries on eligible capacity errors.

One detail is easy to miss. Prompt caches created by 2h and 8h requests cannot be reused by ordinary Standard requests, even if the prompt is identical word for word. An application that sends half its traffic through discounted windows and half in real time will see its cache hit rate split in two.

Multi-step agent tasks can also get the discount. Follow-up requests in the same workflow must keep the same API project, the same model, the same service tier, the same background setting, the same completion window and the same continuation_id. The application generates this ID itself, and reusing it also earns higher queue priority. If the previous response has expired, the original input and all output items must be replayed in order, including encrypted reasoning items and the raw function calls.

Flex is on its way out, Batch stays

One change in the documentation carries real weight: starting with GPT-6.2 Astra and Sol, new models will no longer offer Flex processing (service_tier: flex) at launch. Long-running requests should use discounted windows with Standard instead. Flex was OpenAI's low-price lane for requests where slower is fine, and that demand is now being absorbed into the two completion windows in background mode.

Set side by side with the Batch API, the division of labor is fairly clear. Batch fits submitting a large set of independent requests at once, and the documentation recommends continuing to use it for that. The 2h and 8h tiers of background mode are aimed at individual long tasks, especially agent runs that call tools and proceed in multiple steps. Results can be collected by webhook, and a run can be canceled midway by calling POST /v1/responses/{id}/cancel; repeated cancellations are idempotent.

A rough calculation: a document-review job that runs once a night and consumes 100 million tokens would see its token cost billed at 60% after moving to the 8-hour tier. Converting that to dollars depends on each model's unit price and the input-to-output ratio. OpenAI has not published a separate price list for each window this time, and per-million-token prices circulating online cannot currently be checked against official pages, so this article does not use them.

Data is stored temporarily on disk

Teams that handle compliance should note one clause: background mode temporarily stores response data so that requests can run to completion and clients can retrieve results, even if store: false is set or the organization has Zero Data Retention (ZDR) enabled. OpenAI says the temporary data for all three windows is encrypted at rest; the documentation does not say how long it is kept.

For healthcare and financial customers that have signed zero-data-retention agreements, this is a genuine trade-off: to get a 40% discount, they must accept that data briefly sits on OpenAI's side. There is a cost in experience too. The documentation warns that the first token of a background response may arrive more slowly than in a synchronous request, and the Java and Ruby examples simply set the timeout to 30 minutes.

Sources: OpenAI developer documentation, Background mode guide, used to verify the 2h and 8h discount rates, eligible models, the Flex change and the temporary-storage terms; CocoLoop; OpenAI Developer Community.