Vibeleaderboard
← All Intel
Intel / post

OpenAI adds 2h and 8h discounted background windows to Responses API

Source
x.com
Date
OpenAIDevs@OpenAIDevs

Not every workflow requires an immediate response. For single- or multi-turn workflows that can wait, save up to 40% per token than Standard on GPT-6.2 Astra or GPT-6.2 Sol. In Responses API background mode, choose a 2h or 8h target for the first response. These discounted windows are now in public beta.

Context

OpenAI now charges less for Responses API requests that can wait hours for an answer. A developer who accepts a two-hour target for the first response pays model rates 30% below Standard, OpenAI's ordinary pay-as-you-go processing, and one who accepts an eight-hour target pays 40% below. The OpenAI Developers post gives only the headline figure of "up to 40%"; the split between the two windows comes from OpenAI's background mode guide and API changelog, which dates the public beta to November 17.

Background mode is a setting in the Responses API, the OpenAI interface that applications use to send prompts to its models and get answers back. Instead of holding a connection open while the model works, the application sends the request, keeps the ID the API returns, and collects the result later, either from a webhook (a message OpenAI sends to the application's server when the job finishes) or by checking the ID. The new option is `background_options.completion_window`. The default, `soonest`, carries no discount. With `2h` or `8h`, the clock starts when the API accepts the request and applies to that first response. OpenAI describes this as a target, saying the API "targets completion" within the window, and the guide does not say what happens when a request runs longer. It does say the API retries eligible capacity errors automatically once it has accepted a request.

The "multi-turn" part of the post refers to workflows. If the first response asks the application to run a tool, such as looking up a document, the application runs it and sends the result back in a follow-up request. Each eligible follow-up gets the same discount, and reusing the same continuation ID, model, window and project makes follow-ups eligible for higher queue priority, though the guide adds that their start times depend on scheduling. The discount comes with conditions: the request must use Standard processing and must not stream its output, it counts against the same rate limit as Batch API jobs, and ordinary requests cannot reuse the prompt cache those discounted requests create. The pricing page and changelog extend the discount to GPT-6.1 Terra and GPT-6 Luna as well as the two GPT-6.2 models the post names.

OpenAI already sold two cheaper, slower options, and the new windows sit beside them. The Batch API takes a file of independent requests, finishes each batch within 24 hours and charges 50% less than the synchronous APIs, and the pricing page lists Batch rates for the GPT-6.2 models too. Flex processing bills individual requests at Batch rates in exchange for slower responses and occasional unavailability, but the background guide says new models, starting with GPT-6.2 Astra and Sol, launch without Flex and points long-running Responses API work on them to the discounted windows instead. On GPT-6.1 Terra and GPT-6 Luna, which still have Flex prices, Flex remains the lower per-token rate, at half the Standard price against 40% off for the eight-hour window. The guide points independent requests that can be submitted together up front to Batch, and describes background mode as suited to multi-step agent runs, in which the application uses one result to decide the next request. The new option therefore matters most for multi-step agent jobs on the GPT-6.2 models that can wait hours, since those models have no Flex tier. The video attached to the post could not be extracted, so anything shown only there was not reviewed.

Terms in this piece · Glossary
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
More from OpenAIDevs
Recommended reads
Comments

Checking sign-in…

Loading comments…