sysnote
News

OpenAI adds cheaper background requests, with the completion target scoped to the initial response

OpenAI has added two-hour and eight-hour completion windows to background requests in the Responses API. Eligible model-token rates fall by 30% and 40% respectively. The window targets the initial response; a tool-using agent may still have work left after that response arrives.

The November 17 API update puts the feature in public beta for GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. It extends background mode, which lets an application submit a request and retrieve the result later.

The first answer may be a tool call

The background guide gives a document-review example: the model first asks the application to fetch a document, then receives it in a follow-up request. The selected completion target starts when the API accepts the initial request and applies to that response. It does not establish when the whole review will finish.

Follow-ups can also receive the discount. Reusing the continuation ID, window, model and API project makes them eligible for higher queue priority; their start times still depend on scheduling.

Discounted requests use background_options.completion_window set to 2h or 8h, with background enabled and Standard processing. Streaming must be off. Ordinary Standard requests cannot reuse prompt-cache entries created by those discounted windows.

What the discount buys

At short-context Standard rates, GPT-6.2 Sol costs $2 per million input tokens and $10 per million output tokens. Applying the published discounts gives:

Completion window Input per million tokens Output per million tokens
Two hours $1.40 $7
Eight hours $1.20 $6

These are model-token rates calculated from the pricing table, rather than a price for an entire agent run. A job's total depends on the tokens it uses and any separately billed tools.

For independent requests that can be prepared together, the Batch API remains cheaper at 50% below Standard, with a 24-hour turnaround. The background workflow can instead continue after a tool result becomes available, without preparing all requests in advance.

OpenAI's data-controls documentation allows the two discounted windows in Zero Data Retention projects, treating store as false. It still temporarily stores application state to process the request and make the response retrievable.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.