Vibeleaderboard
← All Intel
Intel / post

Responses API now ties reasoning items to prior conversation history

Source
x.com
Date
OpenAIDevs@OpenAIDevs

Agents using GPT-6.2 in the Responses API can build on earlier reasoning to deliver better results with fewer tokens and more cache hits. Starting with our newest frontier models, each reasoning item is now tied to the conversation history before it, so editing earlier messages causes the API to drop reasoning that follows the change.

Context

OpenAI's developer account says that agents using GPT-6.2 in the Responses API can build on their earlier reasoning, which it says brings better results with fewer tokens and more cache hits. The condition comes in the post's second sentence: each reasoning item is now tied to the conversation history before it, so editing earlier messages makes the API drop the reasoning that follows the edit. A reasoning item is what the API returns alongside a 's answer, a record of the model's thinking on that turn that the application can send back but cannot read. According to OpenAI's reasoning guide, when a later request carries earlier response items, by pointing to a previous response, using a stored conversation, or resending the full history, supported models can use compatible reasoning from earlier turns instead of working it out again. The guide says reasoning tokens are billed as output tokens, which is where the claimed token savings would come from.

OpenAI's guide to keeping the prefix unchanged explains the check. A reasoning item's prefix is the conversation history that comes before it, and for "newer frontier models" the API drops reasoning that follows a change to that prefix. The reasoning guide is more specific: on GPT-6.2 and later, inserting or editing content before a reasoning item removes it, while deleting whole earlier items or appending new ones after it does not by itself. The changes OpenAI lists include routine agent housekeeping: shortening an old tool result, rewriting a developer message, replacing older turns with a summary, or changing request settings such as the tools or the output format. The prefix guide says the API drops the reasoning "as part of our safeguards," and the reasoning guide groups the rule with other removals, such as switching to a less capable model or another organization, that it calls safeguards against , the training of one model on another's outputs. Neither page says how an edited prefix relates to that risk. A field in the response, `input_transformations`, reports which reasoning items were removed and why. The guide to using GPT-6.2 adds that certain model changes, including a downgrade, and a change of end user or organization also drop reasoning, and warns that answer quality can decline and the model may use more tokens repeating earlier work. That is the rule described in coverage of the GPT-6.2 developer launch on November 17. The new post presents the same behavior as a benefit for agents that leave their history alone.

The cache claim follows from how works. Caching matches a request from its start, so a change can cut reuse from that point on, and the same guide calls keeping the prefix intact a long-standing best practice for caching. One habit therefore keeps both the reasoning and the cheaper cached rate, which for GPT-6.2 Astra is $0.40 per million input tokens against $8 uncached, according to its model page and as Simon Willison noted at launch. The guide's advice to developers who need to change something is to append rather than edit: add a `configuration_update` item to change reasoning effort or output format, append a new developer message instead of rewriting an earlier one, add tools with an `additional_tools` item, and shorten long tool outputs or images before they are first sent rather than afterward. For conversations that grow too long, OpenAI's own , through `/responses/compact` or `context_management`, keeps later reasoning items even though it changes the prefix. A custom summary should go at the end of the context, or the reasoning that follows it may be dropped.

The quality and efficiency gains are OpenAI's claims. The post gives no figures, and nothing read here measures how much results or token use change when reasoning is kept rather than dropped. The documents tie the rule to GPT-6.2 and later and to "newer frontier models" without listing which models those are, and the video attached to the post could not be extracted, so anything shown only there was not reviewed. What the documents do establish is the practical rule: on GPT-6.2, an agent should treat its history as append-only, and one that shortens, rewrites or summarizes earlier turns in place should check `input_transformations` to see what reasoning it is losing.

Terms in this piece · Glossary
  • reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
  • distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
  • prompt caching — Reusing the model's processed form of a repeated prompt prefix so subsequent calls skip re-reading it, cutting cost and latency substantially.
  • context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
More from OpenAIDevs
Recommended reads
Comments

Checking sign-in…

Loading comments…