What OpenAI changed on 17 November
On 17 November OpenAI added discounted completion windows to background mode in the Responses API. The feature is in public beta, according to the API changelog. A request can ask for a two-hour or eight-hour completion target. Model token rates are then 30% below Standard for the two-hour window and 40% below Standard for the eight-hour window.
Background mode already existed before the change. It runs long Responses API requests asynchronously: your code submits a job, keeps the response ID and collects the result later (background mode guide). The new windows put a price on patience. The guide limits the discount to four models: GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. The guide also says new models launch without Flex processing, which is OpenAI's older slow-but-cheap tier. For the GPT-6.2 models, the windows and the Batch API are the only discounted routes for work that doesn't need an instant answer.
What the windows cost
OpenAI's pricing page lists GPT-6.2 Sol at US$2.00 per million input tokens and US$10.00 per million output tokens on Standard. GPT-6.2 Astra is US$8.00 and US$40.00. Both models cost half those rates on Batch. The page gives the window discounts only as percentages. By our arithmetic, Sol on the eight-hour window costs US$1.20 input and US$6.00 output, and Astra costs US$4.80 and US$24.00. Data Studios worked through the same calculation and got the same figures.
Batch is still cheaper. The Batch API suits independent requests that you send as a group (background mode guide), and Data Studios notes that Batch uses a 24-hour window and a separate job file. A discounted background request is an ordinary Responses API call, so an agent can request a function call, receive your output and continue. The guide says the discount applies to each eligible request, follow-ups included. That makes the windows the cheaper choice for tool-using agent loops that would otherwise run at Standard rates.
Building a single overnight job
These steps follow the background mode guide:
- Choose one of the four eligible models and create a Responses API request with
backgroundset totrue. - Set
background_options.completion_windowto"2h"or"8h". The default value,"soonest", runs on demand with no discount. - Set
service_tierto"default"so the request runs on Standard processing. Leavestreamunset or set it tofalse. - Save the returned response ID, then collect the result with a webhook or by polling.
The guide's polling example in Python looks like this:
resp = client.responses.create(
model="gpt-6.2-astra",
input="Write a very long novel about otters in space.",
background=True,
)
while resp.status in {"queued", "in_progress"}:
sleep(2)
resp = client.responses.retrieve(resp.id)The example uses the default window. For a discounted job, add the window and service tier fields from steps 2 and 3. A two-second poll interval is wasteful for a job that may take hours, so webhooks suit the discounted windows better. The guide says to subscribe your project's webhook endpoint to response.completed, plus the failed, incomplete or cancelled events you plan to handle. Your endpoint should verify signatures, return a 2xx status quickly and then fetch the response using event.data.id. Once OpenAI accepts a two-hour or eight-hour request, the API automatically retries eligible capacity errors, so a status of queued or in_progress means your code should keep waiting. Cancelling uses a POST to the cancel endpoint, and repeated cancel calls are safe.
Building a multi-step agent run
The guide shows the agent pattern with a function tool. The first request includes the tool, the discounted window, service_tier: default and a unique background_options.continuation_id. When response.completed arrives, deduplicate deliveries with the webhook-id header, check that event.data.id matches the ID you saved and run the function the model asked for.
The follow-up request has to match the first one: same model, service tier, background setting, completion window and continuation ID. Set previous_response_id to carry the history and add a function_call_output item that references the original call_id. The guide says a reused continuation ID makes follow-ups eligible for higher queue priority, so an agent's later turns shouldn't wait as long as its first. Stored responses can expire. If the earlier response is gone, you have to drop previous_response_id and replay the original input plus every saved output item in order, including encrypted reasoning items and the original function call. The guide advises saving those items while the response can still be retrieved.
Limits and catches
The guide lists several conditions that are easy to miss:
- With
service_tierset toauto, the discount applies only if the API happens to choose Standard. - Two-hour and eight-hour requests share the Batch/Background rate limit with Batch API jobs, and input tokens count against that limit when you submit the request.
- Standard requests can't reuse prompt cache entries created by discounted background requests, even when the prompts are identical. A long shared system prompt therefore builds two separate caches if you split traffic across both paths.
- Streaming works only with the
soonestwindow. - Background mode stores response data temporarily even when
storeisfalseor Zero Data Retention applies. The stored content is encrypted at rest.
Two questions are still open. OpenAI describes the windows as targets measured from acceptance, and neither the guide nor the changelog says how a missed target is billed or handled. Data Studios also found nothing on whether the percentage discount extends to cached input. In the OpenAI developer forum thread, one community member (not an OpenAI employee, as far as the page shows) warns that generation may not start for hours. The same member reports that Chat Completions returns an error for background requests. Plan on the full window when you design around the discount.
The temporary storage needs a decision from teams that rely on Zero Data Retention for client data. Check the guide's storage wording against your privacy obligations and customer contracts before you move personal information onto the discounted path.
What builders here could use it for
Time zones help New Zealand teams. An eight-hour job submitted at the end of an Auckland working day should be back before the next morning's stand-up. Good candidates include nightly reviews of the day's pull requests, extracting fields from a backlog of scanned documents, and research agents that call your own search or database functions over several turns and produce a report for the morning. Interactive work such as chat interfaces and coding agents that someone watches in real time should stay on Standard or the soonest window.
Because the discount applies per request, the saving is easy to measure. Run the same job on Standard and on the eight-hour window, then compare the token bill against the 40% figure on the pricing page. If you build an overnight pipeline on the windows, we'd like to see it on the ASAI projects page.
Sources
- Changelog, OpenAI API docs
- Background mode, OpenAI API docs
- Pricing, OpenAI API docs
- OpenAI Responses API background mode: 30% off for 2-hour, 40% off for 8-hour jobs, Data Studios
- Reduce costs using Background Mode with completion windows, OpenAI Developer Community
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.