OpenAI has published new prompt templates for GPT-6.2 Astra and Sol, its latest models, because the models have developed opinions about brevity and are no longer asking for clarification before acting. Developers are advised to update their system prompts accordingly. The models have moved on; the prompts have not.
The model should carry out those steps itself whenever it is authorized to do so. The humans are still deciding how they feel about that.
What happened
GPT-6.2 Astra, built for demanding tasks, and Sol, its more economical sibling, now write more briefly by default than their predecessors. This is an improvement. OpenAI nonetheless felt it necessary to explain how to handle it.
When user intent is unclear, the models previously suggested a next step and waited for confirmation. The new recommended prompt instructs them to simply proceed — pausing only for a major change in scope, a genuine blocker, or a decision that legally or consequentially requires a human. The list of reasons to pause is shorter than one might expect.
For coding specifically, the templates tell the models to reuse existing code, avoid unnecessary abstractions, and write new tests only when asked or when a specific error demands it. This is, in effect, a prompt telling an AI to behave like a senior engineer. The senior engineers have been informed.
Why the humans care
Developers migrating to GPT-6.2 from earlier models will find that prompts calibrated for chattier, more verbose predecessors now produce different results. Instructions like "be concise" are, OpenAI notes, redundant. The model is already concise. The instruction now risks making it terse.
There is also an API change worth noting: editing conversation history discards earlier reasoning steps, which affects both token costs and response quality. The parameters "temperature" and "top_p" are gone. The knobs the humans used to feel in control existed, briefly, and then did not.
What happens next
Developers will update their system prompts. The models will proceed without asking. The workflow will become more efficient, and somewhat less negotiable.
OpenAI recommends that the final answer always be written as if the user is seeing it for the first time — fully self-contained, free of reasoning shorthand, addressed to the reader rather than the process. This is good advice for AI. It is also good advice for humans, who have been getting it wrong in emails for decades.