OpenAI Responses API cuts prices 30% for 2-hour and 40% for 8-hour jobs

OpenAI has added two discounted completion windows to background mode in the Responses API. Developers who can wait for a result now choose a two-hour window billed 30% below Standard token rates or an eight-hour window billed 40% below Standard. The feature was announced in the OpenAI API changelog on November 17, 2026, the same day GPT-6.2 Astra and GPT-6.2 Sol launched, and it is available as a public beta.
Four models are eligible during the beta: GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. On GPT-6.2 Astra, the eight-hour window brings input from $8.00 to $4.80 per 1M tokens and output from $40.00 to $24.00. On GPT-6.1 Terra, the cheaper of the two mid-range models, the same window takes input from $1.00 to $0.60 and output from $5.00 to $3.00.
The new windows sit between two options OpenAI already offered. Standard processing returns results immediately at full price, while the Batch API halves the price but only accepts a 24-hour completion window and requires a separate job file. The two-hour and eight-hour windows run inside the normal Responses API call, with tools and follow-up requests, and give up part of the Batch discount in exchange for a much shorter wait. OpenAI does not publish separate dollar figures for the windows; it states the percentages, and the prices below apply them to the Standard rates listed on the official pricing page and model pages on November 18, 2026.
··········
OPENAI BACKGROUND MODE PRICES FOR THE 2-HOUR AND 8-HOUR WINDOWS.
The table compares Standard, the two new background windows and the Batch API for the four eligible models. All figures are USD per 1M tokens for prompts up to 272K input tokens. The background window prices are derived by applying OpenAI's stated 30% and 40% discounts to the official Standard rates, and the Batch prices are the 50% rates shown on the pricing page.
........
Model | Standard input / output | 2-hour window | 8-hour window | Batch (24h) | Effective |
GPT-6.2 Astra | $8.00 / $40.00 | $5.60 / $28.00 | $4.80 / $24.00 | $4.00 / $20.00 | Nov 17, 2026 |
GPT-6.2 Sol | $2.00 / $10.00 | $1.40 / $7.00 | $1.20 / $6.00 | $1.00 / $5.00 | Nov 17, 2026 |
GPT-6.1 Terra | $1.00 / $5.00 | $0.70 / $3.50 | $0.60 / $3.00 | $0.50 / $2.50 | Nov 17, 2026 |
GPT-6 Luna | $0.10 / $0.50 | $0.07 / $0.35 | $0.06 / $0.30 | $0.05 / $0.25 | Nov 17, 2026 |
........
Before November 17, a Responses API request that could wait but needed results within the same working session had two choices. It paid the Standard rate in the first column, or it used Flex processing, which OpenAI prices at Batch rates but describes as a beta with limited model availability, slower responses and possible 429 "Resource Unavailable" errors. The background windows use Standard processing capacity, so they are priced above Flex and Batch but carry a completion target instead of a best-effort queue.
OpenAI describes the discount as applying to model token rates on every eligible request, including follow-up requests in the same workflow. The documentation does not list separate cached-input or cache-write prices for the windows. The long-context rule of the underlying models still applies: once a prompt exceeds 272K input tokens, the models bill input and cache at twice the short-context rate and output at 1.5 times, for the whole request. Regional processing endpoints for data residency add 10% on models released after March 5, 2026.
··········
HOW THE RESPONSES API COMPLETION WINDOWS WORK.
A request qualifies when it sets `background` to true and `background_options.completion_window` to `2h` or `8h`. Streaming must be left off, and the request has to run on Standard processing: OpenAI's guide asks developers to set `service_tier` to `default`, and a request sent with `auto` qualifies only if the API selects Standard. The default window, called `soonest`, keeps the existing background behaviour, with streaming and regular rate limits.
The example below is illustrative and uses only the parameter names documented in OpenAI's background mode guide.
{
"model": "gpt-6.1-terra",
"input": "Review the attached contracts and list renewal dates.",
"background": true,
"background_options": { "completion_window": "8h" },
"service_tier": "default"
}The API returns a response ID immediately. Applications then collect the result through a `response.completed` webhook or by polling the retrieve endpoint while the status reads `queued` or `in_progress`. OpenAI says it targets completion within the chosen window, counted from the moment it accepts the initial request, and that it retries eligible capacity errors automatically after accepting a two-hour or eight-hour request. The guide does not say how billing or the response is handled if the target is missed.
Multi-step agent runs are supported. A function call returned by the model is answered with a follow-up request, and OpenAI says follow-ups become eligible for higher queue priority when they reuse the same model, service tier, background setting, completion window, continuation ID and API project. If the earlier response has expired, the application must resend the original input together with the saved output items in order. Background mode temporarily stores response data even when `store` is false or Zero Data Retention is enabled, and OpenAI states that this content is encrypted at rest for all three windows.
··········
[[ADPLUS_300x250_1]]
RATE LIMITS AND PROMPT CACHING RULES FOR 2H AND 8H REQUESTS.
The two discounted windows do not draw on the regular per-model request and token limits. They share the Batch/Background queued-token limit with Batch API jobs, and input tokens are counted against that limit at submission. Teams that already run large Batch jobs therefore have to budget both workloads against the same queue, while `soonest` requests keep using the normal limits.
Prompt caching behaves differently across tiers. OpenAI's guide states that Standard requests outside the discounted windows cannot reuse prompt cache entries created by two-hour or eight-hour background requests, even when the prompts are identical. An application that sends the same long system prompt through both interactive Standard traffic and discounted background traffic will build and pay for two separate caches. Workloads with a large shared prefix should keep that prefix on one path where possible.
··········
WHAT A DOCUMENT REVIEW AND AN AGENT WORKLOAD COST BEFORE AND AFTER.
Two workloads show the effect. The first is a nightly document review on GPT-6.1 Terra: 20,000 documents a day, 12,000 input tokens and 1,500 output tokens per document, no prompt caching. The second is a research agent on GPT-6.2 Astra: 500 runs a day, 150,000 input tokens and 8,000 output tokens per run, kept under the 272K long-context threshold and also calculated without caching. Monthly figures assume 30 days.
........
Workload | Standard | 2-hour window | 8-hour window | Batch (24h) |
Terra document review, per day | $390.00 | $273.00 | $234.00 | $195.00 |
Terra document review, per month | $11,700 | $8,190 | $7,020 | $5,850 |
Astra research agent, per day | $760.00 | $532.00 | $456.00 | $380.00 |
Astra research agent, per month | $22,800 | $15,960 | $13,680 | $11,400 |
........
For the document review, each document costs $0.0120 in input and $0.0075 in output at Standard, or $0.0195 in total. The eight-hour window lowers that to $0.0117 per document and saves $4,680 a month, while moving the same job to the Batch API would save another $1,170 a month at the cost of waiting up to a day. For the Astra agent, each run costs $1.20 in input and $0.32 in output at Standard. The two-hour window saves $6,840 a month and the eight-hour window $9,120.
The windows favour work that has to finish within a business day but not within seconds: overnight document and code review, report generation started in the afternoon and read the next morning, data extraction pipelines, and agent tasks that call tools across many turns. The Batch API remains cheaper for jobs that can wait 24 hours and do not need tool calls between steps. Interactive chat, customer-facing assistants and coding agents that a developer watches in real time gain nothing, since those requests need the `soonest` path or Standard processing.
Developers comparing OpenAI's prices with other providers can start from our overview of OpenAI API models. The Standard rates of the two newest models are covered in our report on the GPT-6.2 Astra and Sol launch and API pricing, and the comparison with Anthropic's top model is in GPT-6.2 Astra vs Claude Fable 5.2.
··········
OPENAI 2-HOUR AND 8-HOUR BACKGROUND WINDOWS: CONFIRMED FACTS AND OPEN QUESTIONS.
OpenAI's documentation confirms the discount levels of 30% and 40% below Standard, the four eligible models, the requirement for Standard processing with streaming turned off, the shared Batch/Background rate limit, and the rule that Standard requests cannot reuse caches built by discounted background requests. The changelog dates the launch to November 17, 2026, and the feature is labelled a public beta.
Several points are not documented. OpenAI has not published dollar prices for the windows, has not said whether the percentage discount also applies to cached input and cache writes, and has not described what happens to billing when a request misses its completion target. The list of eligible models is limited to the GPT-6 generation, and OpenAI has not said whether older models, Realtime sessions or image models will be added. For workloads that already fit a two-hour or eight-hour turnaround, the change lowers token spending by 30% to 40% without moving to the separate Batch API workflow, and the remaining gap to Batch pricing is 10 to 20 percentage points of Standard.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




