Back Home

推論系統

Migrating to GPT-6.2: Rewriting Older Conversations May Make the Model Repeat Its Reasoning

For multi-turn assistants and agents, developers should review summarization, tool-output truncation, and instruction-update flows in addition to retesting prompts. The official compaction mechanism has an exception that preserves reasoning items.

zh-Hant

When migrating multi-turn assistants or agents to GPT-6.2 Astra or Sol, developers should check whether their applications rewrite content the model has already read. OpenAI’s documentation states that, starting with GPT-6.2, editing conversation history, switching certain models (including downgrading), or removing reasoning items after an end user or organization changes may reduce response quality and increase token usage as the model repeats previous work. Official developer guidance

Both models are available through the API. The OpenAI announcement does not specify a release date for the developer guidance. Search Engine Journal covered the prompting recommendations and history-editing rules on November 18, 2026, but that confirms only that the guidance had been reported by then; it does not establish that the rules first went live that day. Original report

Saving context may come at the cost of recomputation

Reasoning items are data the model can use to continue earlier reasoning. The conversation content that precedes an item is called its “prefix.” OpenAI lists shortening old tool results, rewriting developer messages, and replacing earlier turns with summaries as examples of prefix changes. Such changes may cause subsequent reasoning items to be removed, leading to extra computation, cost, and latency. Reduced quality is also a possible consequence, though it does not necessarily happen with every edit. Prefix management documentation

Suppose an auditing agent reads a statement, and the application later replaces that old statement with a summary to save context before sending the conversation back to the model. From the user’s perspective, the task is still ongoing, but the model’s conversation history has changed. This hypothetical example follows from the documentation; it is not a failure observed in testing. During migration, developers should check for operations like this that may be hidden in conversation-management code.

Handle instruction updates and history compaction separately

OpenAI recommends updating instructions by appending a new developer message at the end. If large tool results need to be truncated, do so before first sending them to the model whenever possible. Integration recommendations

Long conversations have a separate compaction exception: when using OpenAI’s compaction mechanism through /responses/compact or context_management, the system preserves subsequent reasoning items even when the prefix changes. Replacing older turns with a custom summary should not be assumed to work the same way. The documentation also distinguishes caching from reasoning continuity: prompt-cache diagnostics can help identify history differences, but they do not prove that the model reused its reasoning. Check input_transformations in the response to confirm which items were removed and why. Compaction exception and diagnostic guidance

Prompt retesting should cover completeness and task completion

On prompting, OpenAI says Astra is more concise and direct by default than earlier models, and recommends reviewing instructions such as “keep answers short” during migration. OpenAI also notes that when user intent is unclear, Astra may suggest a next step and wait for confirmation. For long tasks, specify the desired outcome and which decisions the model may make on its own. These are vendor observations and prompting recommendations, not independent measurements by this article. Prompting guidance

Based on these rules, this article recommends splitting acceptance testing into two groups. In one, compare the original and revised prompts on the same tasks, checking whether answers retain necessary context and complete the authorized work. In the other, compare flows such as normal continuation, instruction updates, custom summarization, and model downgrades, recording quality, token usage, latency, and the reasons reasoning items were removed. This test design is derived from the documentation; the sources provide no cost increase that can be applied directly.

Keep a record of the messages actually sent, not just the text shown in the chat interface. That makes it possible to trace differences in results to the specific steps by which the application changed the conversation history.

Sources

  1. Using GPT-6.2 | OpenAI API
  2. Keep the prefix unchanged | OpenAI API
  3. Introducing GPT-6.2 Astra and Sol
  4. OpenAI Publishes Prompting Guidance For GPT-6.2 Astra