OpenAI recommends that developers review their brevity guidelines for GPT-6.2 Astra, which launched on November 17 with GPT-6.2 Sol. The company describes Astra as more direct by default than its predecessors. The migration also includes an announced change in the preservation of reasoning by the API.
Modifying the history could erase the reasoning
OpenAI announces that starting with GPT-6.2, a change to history by the client detected by the API will cause reasoning elements to be aborted. The same treatment is expected after a downgrade or change of end user account or organization. These elements contain the reasoning from previous exchanges. Depending on the company, deleting them may deteriorate responses and require more tokens to redo work already done.
Examples given by OpenAI include a shortened tool result, old exchanges replaced by a summary, and a rewritten developer message. The latter carries the instructions for the application. To make them evolve, the company recommends adding a new developer message which asks to apply the new instructions while ignoring the previous ones, without retouching the original. Some editorial tools can modify their instructions themselves during the conversation: this practice could therefore cause them to lose previous reasoning and lead to an additional expenditure of tokens.
Astra can wait for a human decision
When the request lacks precision, GPT-6.2 Astra can propose the continuation of the work then wait for an agreement, indicates OpenAI. For tasks to be carried out independently, the company advises explicitly defining the expected result and the decisions that are up to the user. The proposed instructions allow reasonable, reversible and explicit hypotheses to continue the work.
The stop must then take place if the scope changes significantly, if an obstacle blocks execution or if a human decision is necessary. A verification with the user is also planned before any destructive or irreversible operation. OpenAI also describes an asynchronous tool: the model can ask a question without interrupting work that does not depend on it. This tool must be defined by the application, which displays the question; it is not provided directly in the responses API.
The evening letterFree · 7 p.m.
Do you code with AI?
Deciphered AI dev tools, agents and new features, every evening in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Adapt texts to the reader and the task
For emails, documents or other writings addressed to an external audience, OpenAI recommends specifying the readership and providing elements of context that they do not know. In revision, the proposed guideline is to retain the original content and wording, subject to changes requested by the user. The company also describes an adjustable tone, from professional colleague to conversational assistant. His example intended for daily work requires taking the user’s vocabulary and its sources, then placing the essentials at the top.
Another precaution concerns responses produced after long tasks. According to OpenAI, they may retain abbreviations, jargon, or shortcuts from middle reasoning or other agents. The company recommends requesting a final stand-alone text, understandable by a person who has not followed the progress reports, without these shortcuts.
Settings to review during migration
Editorial tools moving from GPT-6 Astra to the new version are particularly affected by the revision of the instructions. OpenAI described the old model as readily detailed, with lists, tables and Markdown intended to make it easier to read. For less formatted prose, the company suggested requiring concise paragraphs and reserving lists for parallel or sequential data. She had previously pointed out that GPT-5.6 was generally less verbose than GPT-5.5 and that some tasks no longer required additional brevity guidelines.
Astra and Sol are accessible by the API identifiers gpt-6.2-astra and gpt-6.2-sol. OpenAI presents its examples as a starting point based on the behaviors observed on these models. The GPT-6 recommendations remain accessible in a tab on the same page as those of GPT-6.2, to compare the indications between versions.
For a new integration, OpenAI recommends a medium reasoning effort. Applications previously set to zero or minimal effort are encouraged to choose the low level.

