GPT-6.1 Terra became available in the Responses API on 26 October 2026. Before moving a long-running workflow to a different model, check how your app continues a conversation after server-side compaction. Confirm the target model supports the request settings you use, and check its current API tier and pricing.
OpenAI documents two ways to continue Responses work after compaction. An application can manage an input array, or carry previous_response_id and submit the new message. Combining snippets from both examples without checking their state model can build the wrong request.
Identify who carries the history
With stateless input arrays, append the returned output items, including compaction items, as the guide describes. With previous_response_id, carry the ID forward and send only the new user message for that turn. The compaction item is opaque; it is not text you should decode and rewrite as a summary.
For a small test, give the app an invented instruction to label each result SAMPLE-17. After observing a compaction item, send one new task and check that the next result still uses that label. Record which continuation method and request shape you used. Use synthetic input for the check.
Change one mechanism at a time
The current guide’s pruning advice differs between the patterns; it explicitly says not to manually prune an ID-based chain. Preserve the prior working implementation before changing it. A smaller payload or successful HTTP response does not by itself prove the assistant continued from the intended state. This article did not execute an API conversation.