OpenAI's Responses API Gets 2-Hour and 8-Hour Background Windows at 30% and 40% Off

OpenAI's Responses API adds 2-hour and 8-hour background windows in public beta, 30% and 40% below Standard. Setup, costs, Batch comparison, doc gaps.

OpenAI added two discounted background completion windows to the Responses API on November 17, in public beta. A request with a two-hour window is billed at model token rates 30% below Standard, and one with an eight-hour window at 40% below. The windows work on gpt-6.2-astra, gpt-6.2-sol, gpt-6.1-terra and gpt-6-luna, and arrived the same day as GPT-6.2 Astra and Sol. API changelog · Background mode guide

Everything below comes from OpenAI’s developer documentation: the changelog, the background mode guide, the API reference, and the pricing, Batch, Flex, rate-limit, data-controls and model pages. The per-token prices for the two windows are our arithmetic from the published Standard rates, because the pricing page describes the discount in one sentence and has no table for it. S5 Labs has not run a request through either window, and we found no independent measurement of completion times.

What a completion window is

Background mode runs a Responses API request asynchronously. You create it with background: true, then poll or wait for a webhook. The new background_options.completion_window field takes one of three values:

WindowDiscount on model token ratesOpenAI’s example use
soonest (default)NoneAn on-demand report while someone waits for the result
2h30% below StandardAnalyzing the morning’s support tickets before an afternoon triage meeting
8h40% below StandardReviewing a batch of documents overnight for the next morning

Background mode guide · API reference

The windows are targets. For 2h and 8h, the guide says, “the API targets completion within the selected window after it accepts the request”, and, in its follow-up steps, “start times depend on scheduling”. None of the pages we read promises completion or says what happens when a target is missed. Background mode guide

How to request one

{
  "model": "gpt-6.1-terra",
  "input": "...",
  "background": true,
  "background_options": { "completion_window": "8h" },
  "service_tier": "default"
}

The guide describes these fields without a full JSON example, so this shape is our reading of the object schema in the API reference. The reference states the conditions in one line: “The 2h and 8h completion windows require background: true, Standard processing, and stream: false or omitted.” API reference

Streaming a background response works only with soonest, and the guide adds that “you can’t stream a response later if it was created without streaming”. For the tier, set service_tier to default. With auto, the request qualifies “only if the API selects Standard”. Flex, Fast mode and Ultrafast are separate service_tier values in the API reference, so the Standard requirement rules out all three, though the guide names none of them. Background mode guide · API reference

Once the API accepts a 2h or 8h request, it retries eligible capacity errors on its own. While the status is queued or in_progress, you wait for a webhook or poll the same response ID. The guide says to subscribe to response.completed and to the events for failed, incomplete or canceled responses, to deduplicate deliveries using the webhook-id header, and to check that event.data.id matches the response ID you saved. A POST to /v1/responses/resp_123/cancel cancels a response, and the cancel call is idempotent. The guide does not say whether a canceled request is billed for work already done. Background mode guide

What it costs

Rates are per million tokens, shown as input / output, for prompts up to 272,000 input tokens. Standard and Batch come from the pricing page. The 2h and 8h columns are Standard multiplied by 0.70 and 0.60.

ModelStandard2h8hBatch
gpt-6.2-astra$8.00 / $40.00$5.60 / $28.00$4.80 / $24.00$4.00 / $20.00
gpt-6.2-sol$2.00 / $10.00$1.40 / $7.00$1.20 / $6.00$1.00 / $5.00
gpt-6.1-terra$1.00 / $5.00$0.70 / $3.50$0.60 / $3.00$0.50 / $2.50
gpt-6-luna$0.10 / $0.50$0.07 / $0.35$0.06 / $0.30$0.05 / $0.25

Pricing page

Take a job with 10 million uncached input tokens and 2 million output tokens on GPT-6.2 Astra:

  • Standard: 10 × $8.00 + 2 × $40.00 = $160
  • 2h: 10 × $5.60 + 2 × $28.00 = $112
  • 8h: 10 × $4.80 + 2 × $24.00 = $96
  • Batch: 10 × $4.00 + 2 × $20.00 = $80

The same job on GPT-6.2 Sol costs $40 at Standard, $28 at 2h, $24 at 8h and $20 through Batch.

Two caveats apply to these figures. First, the changelog and pricing page apply the discount to “model token rates” and stop there. They don’t say whether it covers cached-input reads, cache writes or the higher rates for prompts above 272,000 input tokens, so the table stops at uncached input and output on short-context prompts. The second concerns the prompt cache: “Standard requests outside the discounted completion windows can’t reuse prompt cache entries created by 2h or 8h background requests, even when the prompt is identical.” If interactive traffic and background jobs share a long prefix, the background jobs will not warm the cache for the interactive side. API changelog · Pricing page · Background mode guide

Against Batch, Flex and Standard

OptionPrice against StandardTimingAvailability on the four models
Standard, including soonest backgroundFull rateSynchronous, or background with no stated targetAll four
2h window30% lowerTargets two hours from acceptanceAll four
8h window40% lowerTargets eight hours from acceptanceAll four
Flex (service_tier: "flex")Batch ratesNo latency target; can return 429 Resource Unavailable, which is not chargedTerra and Luna, not the 6.2 models
Batch50% lower“within 24 hours (and often more quickly)”All four

Batch guide · Flex processing guide · Pricing page

Batch is still the cheapest route, at 50% off against 40% for the 8h window. The windows give you an ordinary Responses call that can use a function tool and take follow-ups; Batch is a one-shot file of up to 50,000 requests and 200 MB. Batch also documents its deadline: unfinished requests are canceled, and completed ones are delivered and billed.

The pricing page lists Flex rates for Terra, Luna and the prior-generation gpt-6-astra and gpt-6.1-sol, but none for the 6.2 models, so on GPT-6.2 Astra and Sol the windows are the only discounted processing option short of Batch. Pricing page

The windows and Batch count against one limit. The guide says: “Requests with 2h or 8h completion windows share the Batch/Background Rate Limit with Batch API jobs. The API counts input tokens at submission.” That limit caps queued input tokens per model. The guides give no figure and point you to Settings > Organization > Limits. The Terra and Luna model pages do list a “Batch queue limit” by usage tier, starting at 3,000,000 for Terra and 20,000,000 for Luna on the Build tier, while the GPT-6.2 Astra and Sol pages have no such column. A large Batch backlog on a model therefore leaves less room for 2h and 8h requests on the same model. Background mode guide · Rate limits · Batch guide · GPT-6.1 Terra · GPT-6 Luna

Follow-ups in an agent run

This section applies to agent loops you run yourself on v1/responses, the only endpoint the changelog entry lists. OpenAI’s managed Agents API is a separate product.

To tie the requests in a multi-step run together, set background_options.continuation_id to “a unique ID your application chooses for this run” and reuse it on every follow-up. The reference calls it an opaque identifier “used to prioritize related background requests” and adds: “It does not add prior history to the input.” The guide’s follow-up uses the same model, service tier, background setting, completion window, continuation ID and API project. The ID affects queue priority, not price: the guide says reuse makes follow-ups “eligible for higher queue priority” but does not explain how. The discount applies to each eligible request, follow-ups included. API reference · Background mode guide

History is carried separately. Set previous_response_id to the completed response’s ID and send a function_call_output item with the original call_id. If the prior response has expired, omit previous_response_id and replay the original input and every item of the saved output in order, including encrypted reasoning items and the original function call, then append the new function_call_output. Save both while they can still be retrieved, because the guide does not say how long that is. Background mode guide

The completion target “starts when the API accepts the initial request and applies to that response”. Read literally, the target covers that one response and not the whole run, so a run with several follow-ups could take longer than one window end to end. The only tool the guide shows with discounted windows is a function tool that retrieves a document by ID. Background mode guide

Projects with Zero Data Retention can use both windows. The API treats store as false and temporarily stores application state to process the request and make the response retrievable. The background guide says temporarily stored request content is encrypted at rest for all three windows. For the default window the data-controls page says response data is kept “for roughly 10 minutes”; for 2h and 8h it gives no duration. Data controls · Background mode guide

What the docs leave out

  • What happens when a 2h or 8h target is missed, and any SLA.
  • Whether the discount covers cached reads, cache writes and long-context rates.
  • The size of the Batch/Background limit for GPT-6.2 Astra and Sol, and the error returned when it is reached.
  • How long a stored response stays usable, and how long 2h and 8h data is retained.
  • Whether structured outputs or tools other than function calling work with the windows.
  • Whether canceled or incomplete requests are billed.
  • A general-availability date, and whether beta pricing will change.

What to move, and what to leave

Move work that nobody is waiting on but that either should land within hours or needs follow-up requests: overnight evals with tool calls, backfills that fetch documents, and agent runs whose result is read the next morning. The case is strongest on GPT-6.2 Astra and Sol, where Flex is not listed.

Single-shot jobs that can wait a day belong on Batch, which costs less and documents what happens at its deadline. Leave at Standard anything with a hard deadline, anything streamed and anything a person is waiting on. Hold cache-heavy and long-context workloads back until an invoice shows how the discount applies to them.

Before committing a pipeline, run a small job at each window. Log the time from acceptance to completion for every response, and compare the billed amount with the arithmetic above.

Sources

Continue reading.

Insight22 min read

GPT-6.2 Astra Drops to $8/$40, Sol Holds at $2/$10 and Cyber Safeguards Tighten

GPT-6.2 Astra costs $8/$40 per million tokens, 20% below GPT-6 Astra; Sol stays at $2/$10. Migration steps, benchmark caveats and tighter cyber refusals.

Insight13 min read

GPT-6.1 Terra Halves Sol's Price for High-Volume Computer Use

GPT-6.1 Terra costs $1/$5 per million tokens, half of GPT-6.1 Sol. Specs, migration limits, vendor benchmarks, Preparedness results and what to test.

Insight21 min read

OpenAI Decisions API: $0.10 per Million Input Tokens and Nothing for the Answer

OpenAI's Decisions API is GA on gpt-6-luna from October 22: $0.10 per million input tokens, no output charge, priced caching. Where it beats Responses.