Back to blog

Pricing

GPT-6.2 Background Processing: How to Save on AI API Token Costs

OleToken Blog · 7 min read · November 20, 2026

Some AI tasks need an answer immediately. Others can finish before an afternoon meeting or the next morning. GPT-6.2 background processing gives developers a way to separate those workloads, using asynchronous execution and, when eligible, discounted completion windows.

The key qualification is simple: background execution alone does not guarantee cheaper inference. OpenAI's official documentation, checked on November 20, 2026, distinguishes the default soonest window from discounted two-hour and eight-hour windows. Understanding that distinction is essential for accurate AI token pricing and reliable cost estimates.

What background processing means

Background mode runs a Responses API request asynchronously. Your application submits the work, saves the response ID, and retrieves the result after completion. It can receive a webhook or poll for status instead of holding an ordinary synchronous request open until the final answer.

This is useful for long document analysis, report generation, and workflows whose next step depends on a model response. Your application still orchestrates tools and follow-up calls; accepting a background request does not mean the provider automatically executes your entire business process.

Asynchronous execution and discounts are separate

Completion window

Intended timing

Model token discount versus Standard

Soonest

On-demand work

None

Two hours

Work that can wait

30%

Eight hours

Overnight or deferred work

40%

OpenAI describes the longer windows as completion targets after the initial request is accepted. They are not evidence of a guarantee that every multi-step application workflow finishes within one window. Tool time and later requests must be planned separately.

Which GPT-6.2 requests qualify?

The documented discounted windows are in public beta on GPT-6.2 Astra and GPT-6.2 Sol, alongside GPT-6.1 Terra and GPT-6 Luna. They require Standard processing. Setting the service tier to default explicitly selects Standard; auto qualifies only when Standard is actually selected.

Set background mode to true and choose the two-hour or eight-hour completion window. For discounted windows, streaming must be omitted or disabled. The default soonest window supports different streaming behavior but has no completion-window discount.

Avoid confusing the feature with Flex

OpenAI states that new models starting with GPT-6.2 Astra and Sol launch without Flex processing. Discounted background windows are the documented alternative for eligible long-running Responses requests. Do not copy an older Flex configuration and assume it provides GPT-6.2 cost optimization.

Check model, endpoint, project access, and current beta conditions before deployment. A gateway that supports ordinary Responses calls may not implement background options, retrieval, cancellation, or webhooks in the same way.

Calculate the savings correctly

GPT-6.2 Sol lists short-context Standard rates of $2 per million uncached input tokens and $10 per million output tokens. Astra lists $8 input and $40 output. Cached reads and writes have separate rates, and long-context pricing can change the base calculation.

For an illustrative Sol request with 20,000 uncached input tokens and 2,000 billable output tokens, Standard token cost is $0.04 plus $0.02, or $0.06. A qualifying two-hour window reduces that to $0.042; an eight-hour window reduces it to $0.036.

Across 10,000 such requests, the totals are $600, $420, and $360 respectively. These calculations assume identical usage and eligible short-context Standard processing. They exclude tools, taxes, retries, and infrastructure, and are not a complete production quote.

Apply discounts to the applicable charges

The documentation describes discounts on model token rates, not a blanket percentage off every service on an invoice. Paid tools or regional uplifts may affect the final bill. Include all billable reasoning in the output assumption rather than counting only the displayed answer.

For GPT-6.2 requests above 272,000 input tokens, input and cache rates double and output rates increase by 1.5 times for the full request. Calculate the applicable base rates first. Do not price an oversized request as short context simply because it runs asynchronously.

Design a reliable asynchronous AI API workflow

Save the response ID immediately

Persist the response ID alongside your application job ID, selected model, completion window, project, and submission time. A network timeout does not prove the provider rejected the work. Reconcile existing jobs before creating duplicate requests.

Represent queued, running, successful, failed, incomplete, and canceled outcomes in your application. A successful submission is not a successful report. Validate the returned content before marking the business task complete.

Prefer verified webhooks or bounded polling

OpenAI recommends receiving completion events through webhooks and retrieving the response using the event's ID. Verify the webhook signature, deduplicate deliveries, and check that the response belongs to the expected job before acting on it.

If webhooks are unavailable, poll the same response ID while it is queued or in progress. Use a bounded schedule and stop at a terminal state. Do not resubmit the prompt each time you want to check whether the original request finished.

Keep side effects idempotent

A duplicated webhook must not create a duplicated email, payment instruction, or database update. Store processing state and use idempotency mechanisms where supported. Validate and authorize downstream actions independently of model text.

After eligible discounted requests are accepted, OpenAI automatically retries certain capacity errors. While the response remains active, continue tracking that response rather than creating competing copies of the job.

Continue workflows across multiple requests

A background response may request a tool call. Your application retrieves the response, runs the authorized tool, and submits the tool result in a follow-up request. The initial completion target applies to that response, not all future reasoning and tool work.

For higher queue priority on eligible follow-ups, the guide describes reusing the same continuation ID, model, completion window, and API project, together with the required background and Standard settings. A continuation ID groups a run; it is not interchangeable with the response ID carrying conversation history.

If an earlier response expires, the guide describes replaying saved conversation items, including required reasoning and function-call items. Plan result retention and recovery before launching long jobs, and follow the current endpoint documentation rather than reconstructing state from visible text alone.

Compare background processing with Batch

Use Batch for independent requests that can be submitted together in advance. Use background Responses when your application needs the result of one response before deciding the next step. These are different workload shapes, even when both tolerate delay.

The GPT-6.2 model pages list Batch at 50% of Standard token rates. Do not assume the Batch discount stacks with a background-window discount. Choose the supported route that matches dependencies, deadlines, and current terms.

Consider cache boundaries

OpenAI states that Standard requests outside discounted windows cannot reuse prompt cache entries created by two-hour or eight-hour background requests. An identical prompt does not override that boundary. Separate cache forecasts by processing route and inspect real usage.

This can matter for applications mixing daytime chat with overnight analysis. A scheduled job may receive a lower token rate while providing no cache benefit to the interactive workload that follows.

Data handling and operational limits

Background mode temporarily stores response data even when storage is disabled or Zero Data Retention is used, as described in OpenAI's data controls. The guide states that temporarily stored request content is encrypted at rest. Review the applicable retention details for your requirements.

Discounted windows share a Batch/Background rate limit with Batch jobs, with input tokens counted at submission. Queue limits and application deadlines therefore need monitoring alongside spend. Default soonest requests use regular request and token limits.

FAQ

Does setting background mode always save money?

No. Soonest background processing has no completion-window discount. Savings require a supported model, eligible Standard processing, and a discounted two-hour or eight-hour window.

Can I stream a discounted background request?

The documented two-hour and eight-hour windows require streaming to be omitted or false. Use the documented soonest behavior when streaming is necessary, and price it accordingly.

Can a unified platform provide the same feature?

Only if it supports the required options and lifecycle endpoints. Verify pricing, retrieval, cancellation, webhook support, and cache behavior through the actual integration.

Next step

Move one delay-tolerant workload into a measured pilot and compare total cost per accepted result. If you use OLEToken as a unified AI API platform, confirm background-window support and billing terms before relying on these direct-provider discounts.

Sources