Back to Flash News

OpenAI documentation: Responses API adds 2h/8h completion windows

OpenAI developer documentation shows that the Responses API backend mode has added completion window options. When the 2-hour window is selected, the model token rate is 30% lower than Standard; when the 8-hour window is selected, it is 40% lower; the default soonest window does not offer a discount. This feature is currently available in public beta on GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, and GPT-6 Luna.

The discounted windows require Standard processing (service_tier set to default), do not support streaming output, and share rate limits with the Batch API. If subsequent requests continue using the same continuation_id, they can obtain higher queue priority. The documentation also discloses that starting with GPT-6.2 Astra and Sol, new models no longer offer Flex processing.

Officially, the 2-hour window is illustrated by analyzing support tickets from the morning for use in an afternoon triage meeting, and the 8-hour window is illustrated by reviewing a batch of documents overnight for use the next morning. The completion target is counted from when the API accepts the initial request and applies to that response; the discount also applies to eligible subsequent requests.

Source: public information

ABAB AI Insight

OpenAI previously used Flex processing (service_tier=flex) to provide Batch-level rates for workloads that could tolerate slower responses and occasional resource unavailability, and this mechanism once covered some older models. This time, ending Flex on the GPT-6.2 series and shifting the discount path to completion windows in the Responses API backend mode indicates that the company is moving asynchronous discounts from flexible degradation to explicit time commitments.

On the capital path, the 2-hour 30% discount and 8-hour 40% discount sit between Standard full price and Batch API’s deeper discounts, allowing developers to pay in tiers according to latency tolerance. Reusing the continuation_id for subsequent requests can obtain higher queue priority, allowing multi-turn tool-calling workflows to still enjoy discounts, but each response resets the completion target individually.

Compared with Flex, the new windows explicitly require Standard processing and prohibit streaming output, while also sharing rate limits with Batch, reducing the uncertainty of being occasionally available but increasing predictable hourly latency. Judgment: OpenAI is repricing asynchronous capacity through explicit completion windows, steering the medium- and low-priority workloads previously covered by Flex into schedulable queues while tightening the Flex path on new models.

Cognitive laws:

  1. Latency commitments are exchanged for rate discounts

  2. Asynchronous paths shift from flexibility to windows

  3. Queue priority is tied to the continuation token