Latest in AI
OpenAI Responses API Background Windows: Pricing, Limits and When to Use Them
A two-hour or eight-hour window can cut model costs for agent jobs, but only the first response runs against that initial clock.
OpenAI's Responses API now offers two discounted background completion windows in public beta. Choose 2h for model token rates 30% below Standard or 8h for 40% below Standard. The useful case is work that can wait but can't be fully specified up front, such as a document review that might ask your app for another file.
What the background windows cost
The pricing page lists four eligible models: GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. The ordinary soonest window still exists and doesn't get this discount. Here's the arithmetic for GPT-6.2 Sol at its listed short-context rates, assuming 100,000 uncached input tokens and 20,000 output tokens:
| Completion window | Input per 1M tokens | Output per 1M tokens | Example model cost |
|---|---|---|---|
Default (soonest) at Standard | $2.00 | $10.00 | $0.40 |
Two hours (2h) | $1.40 | $7.00 | $0.28 |
Eight hours (8h) | $1.20 | $6.00 | $0.24 |
That's a model-only estimate, not a job quote. A second model request adds another set of tokens. Separately billed tool calls and infrastructure costs are outside the example. When comparing the windows on your own traffic, count the cost of the completed job rather than multiplying the first request by 0.6 and assuming you're done.
How to submit a discounted background request
Use POST /v1/responses and set the processing tier explicitly. Here's an example body for a text-only approval review; in production, include the relevant text in the request. The API reference documents the settings.
{
"model": "gpt-6.2-sol",
"input": "Review the approval notes below for missing sign-offs: [notes]",
"background": true,
"service_tier": "default",
"background_options": {
"completion_window": "8h"
}
}
Keep streaming off. Save the response ID before handing the job off to a worker. Get the result through a verified webhook or by polling that same ID. If it's still queued or in progress, sending the same prompt again creates another request rather than checking on the first one. OpenAI handles eligible capacity retries after acceptance.
What happens when the model needs another file
Suppose the model sees a reference to an approval email that isn't in the supplied notes. If you provide a function to fetch it, the first response can end with that function call. Your app fetches the email, then submits the result in a new response request. OpenAI's target starts at acceptance of the initial request and applies to that response. Don't present an eight-hour model window as an eight-hour delivery promise for the entire workflow.
For related requests, use the same model, project, tier and completion settings. Set a unique background_options.continuation_id for the run starting with the first request and reuse it so later turns can qualify for higher queue priority. The ID doesn't carry conversation history. Pass the prior response ID when available, or preserve and replay state yourself as described in the background guide.
Use Batch for jobs you can list up front
I'd put 5,000 independent classification requests into Batch if a 24-hour turnaround works. You already know each input and can submit them together. I'd use a background window for the approval review because the next input depends on what the first response finds. Keep ordinary processing for a workflow where a person is waiting for the result.
Discounted background requests and Batch jobs share a rate-limit pool, and discounted requests count input tokens at submission. A big overnight Batch run can therefore compete with your review work. Also, ordinary Standard traffic can't reuse prompt cache entries created by discounted background requests. I'd measure a separate token and latency baseline before moving a workflow that already benefits from a warm cache.
Check retention before moving sensitive jobs
A discounted request can require temporary application storage even if store is false. OpenAI's data controls explain how the two discounted windows work with Zero Data Retention projects. If your own policy bans temporary storage, check that requirement before switching; the project setting alone isn't enough to answer it.
Start with a queue you already have, where nobody opens a page expecting the answer immediately. Measure how often the whole job finishes before the person needs it. That's a better test of an eight-hour window than whether one request comes back cheap.
I’m Paras. I write about AI engineering and how teams use it in real products.
I help AI founders explain technical releases like this in terms developers can actually use.