OpenAI Cuts GPT-6.2 Astra API Costs 40% for Overnight Background Jobs
OpenAI's Responses API background mode adds discounted 2h and 8h completion windows, now in public beta. The 2h window costs 30% less than Standard and the 8h window costs 40% less, follow-up requests included. Works on GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna.

Takeaways
- OpenAI's Responses API background mode adds discounted 2h and 8h completion windows, now in public beta.
- The 2h window costs 30% less than Standard and the 8h window costs 40% less, follow-up requests included.
- Works on GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. Sol at 8h costs $1.20/$6 per 1M tokens.
- Reusing a continuation ID gives multi-turn agent follow-ups higher queue priority, which Batch can't do.
- Gotchas: no streaming, Standard tier required, rate limits shared with Batch, prompt caches kept separate.
- Windows are targets, not guarantees. Critics want clarity on missed targets. See pricing .
OpenAI has added two discounted completion windows to background mode in the Responses API, now in public beta. Requests assigned to 2h cost 30% less than Standard processing, while 8h requests cost 40% less. The option targets work that can finish later, such as turning meeting notes submitted at the end of the day into a morning brief.
Background mode runs model requests asynchronously, allowing an application to submit work and retrieve the result later. The beta supports GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, and GPT-6 Luna. OpenAI says model behavior and output quality remain unchanged, although requests may spend longer in the queue.
Three windows, two discounts
The background mode guide defines three completion windows. The default, soonest , uses Standard pricing and suits reports or responses awaiting immediate review. The 2h window fits work such as analyzing morning support tickets before an afternoon triage meeting. The 8h window covers overnight document review, summaries, and scheduled research.
OpenAI starts the timer when the API accepts a request. Each window is a completion target, and delivery can occur after its stated deadline. The following examples come from the pricing page .
| Model | Standardinput / output | 2hinput / output | 8hinput / output | Batchinput / output |
|---|---|---|---|---|
| GPT-6.2 Astra | $8 / $40 | $5.60 / $28 | $4.80 / $24 | $4 / $20 |
| GPT-6.2 Sol | $2 / $10 | $1.40 / $7 | $1.20 / $6 | $1 / $5 |
Batch offers a larger 50% discount for independent requests that can be prepared and submitted together. Background windows cover workflows whose next prompt depends on an earlier model response, tool result, or external event.
Agents get room to loop
A Batch job requires a complete set of independent requests at submission time. An agent may instead request a document, inspect the result, call a tool, and use that tool output to decide its next step. The discounted background window applies to every eligible request in that sequence, including follow-ups.
A continuation ID connects those requests. The client supplies one string for a workflow run and reuses it with the same model, project, and completion window. Eligible follow-ups then receive higher queue priority, reducing the risk that every step in a multi-stage workflow waits at the back of the queue.
Overnight loop, step by step
- Send an
8hrequest with a function tool and a client-generated continuation ID. - Wait for a
response.completedevent through webhooks , or poll with the response ID. - Run the tool call requested by the model.
- Return the tool result with the same window and continuation ID, along with
previous_response_id.
Send an eight-hour request
Queues, caches, and retention
- Streaming: Leave
streamunset or set it tofalse. Streaming supports only thesoonestwindow. - Service tier: Discounted windows require Standard processing. With
service_tier="auto", the discount applies only when the API selects Standard. - Rate limits: These requests share the Batch/Background rate limit with Batch jobs. Input tokens count against the limit when the request is submitted.
- Prompt caches: Standard requests cannot reuse cache entries created by
2hor8hrequests, even when the prompts match. - Temporary storage: Background mode briefly stores response data when
storeisfalseand under Zero Data Retention. OpenAI says the data is encrypted at rest. - Model allowlist: Support is limited to the announced models. GPT-6.1 Sol falls outside that list, and a reported request returned an HTTP 400 error for an unsupported completion window.
- Capacity retries: After accepting a
2hor8hrequest, the API automatically retries eligible capacity errors.
Missed deadlines and billing gaps
The pricing model turns latency into a scheduling input. When a job can finish at any point before a morning deadline, OpenAI can queue it alongside similar work or route it to otherwise idle hardware. The discount compensates developers for that scheduling flexibility.
Joshua Wendt of S5 Labs noted that OpenAI describes the windows as targets and has not documented what happens when a target is missed. He also identified unresolved billing questions involving some prompt-caching and long-context cases.
Early implementation work includes a GitHub proposal that would let users assign 2h or 8h windows to individual agent roles, including overnight proposal generation. A separate pricing analysis noted that GPT-6.2 Sol has no Flex tier, leaving Batch and discounted background windows as its lower-cost processing options.
Jobs that can wait
| Workload | Option | Reason |
|---|---|---|
| Chat replies, fraud checks, and live copilots | soonest | A user or system is waiting for the result. |
| Support-ticket analysis before a scheduled meeting | 2h | The task has a near-term deadline and can tolerate queueing. |
| Nightly summaries, compliance reviews, and morning briefs | 8h | The result remains useful after an overnight run. |
| Multi-step research or reconciliation agents | 2h or 8h | Later steps depend on earlier model responses or tool calls. |
| Large sets of independent prompts | Batch | The requests can be assembled in advance and qualify for the larger discount. |
At Astra rates, a workload using 10 million input tokens and 2 million output tokens costs $160 with Standard processing and $96 with the 8h window, a $64 saving. That pricing makes delayed background processing useful for stateful workflows whose deadlines allow several hours of queueing.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report

