AI NewsModels & agentsAnnouncement
OpenAI tells teams to drop old prompt tricks for GPT-6.2
OpenAI's migration guide for GPT-6.2 Astra and Sol asks teams to drop old prompt tricks. It tells them to delete four sampling parameters when they switch.

Why it mattersMost teams carry prompts and parameters from older GPT models. OpenAI is now saying those will quietly cost them quality and tokens on the new ones.
OpenAI published a migration guide on November 17 for GPT-6.2 Astra and GPT-6.2 Sol. Astra is the new frontier model. Sol is the cheaper one. The guide is on developers.openai.com. It reads as advice for teams moving prompts from GPT-5. The short form: the new models punish habits the old ones rewarded. OpenAI published it on launch day.
What OpenAI says changed under the hood
Two of the changes are behavioural and apply at the API level, not in the prompt. First, OpenAI writes that Astra's output is "more direct and brief by default than prior GPT models". Second, the Responses API now drops hidden reasoning items when a client edits earlier turns in the chat. OpenAI warns this "can affect answer quality and increase token usage". An agent harness that trims or rewrites chat history before a call will trigger that.
A third change is new default behaviour. If the user's intent is unclear, Astra may propose next steps and wait for confirmation instead of running. OpenAI tells teams to spell out what counts as a finished answer and which decisions a user owns.
What OpenAI tells teams to delete or move
The migration quickstart is a short list. Set the model to gpt-6.2-astra or gpt-6.2-sol. Start the reasoning effort at medium, or low if you were using none. Omit temperature, top_p, logprobs and top_logprobs. The models ignore them. OpenAI also says to move stable instructions out of the user turn. Put them in developer messages so they qualify for prompt caching.
The guide also tells teams to prune old lines. "Be concise" fights the new default. Long style rules for lists and tables are no longer needed. OpenAI writes that Astra "tends to use lists, tables and Markdown to make responses scannable". For code, OpenAI asks teams to favour readable code that reuses existing logic. For frontend tasks, it asks for realistic sample images and the removal of "decorative numbering, filler badges".
Where the guide is thin
OpenAI does not state how much any of these changes move cost or quality on a real task. There is no "this saves X percent of tokens" figure, no first-party benchmark, no example diff. The guide lists the knobs and asks teams to measure. OpenAI also does not say which cached items break when a client edits history. A team will have to measure that too.
What a team should do this month
Two cheap checks for a team that already uses the Responses API. Grep the last 100 production prompts for "be concise". Grep them for lists and tables style rules. Grep them for the four deleted parameters. Keep what the model now ignores. The parts that argued with the new default drop out. Then record token cost per completed task before and after the switch, not cost per call. A model that clarifies first before running will spend fewer tokens on a correct answer. It will also spend more tokens on back-and-forth when the brief was vague.
Source
Latest models: migration and prompting, OpenAI developer docs.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


