OpenAI has updated its system-prompt templates for GPT-6.2 Astra and Sol, instructing developers to stop using redundant shortening instructions and deprecated sampling parameters like temperature and top_p. The guidance also highlights that the Responses API now silently discards reasoning content when conversation history is edited, requiring teams to adjust their integrations to avoid increased token costs and degraded quality.

OpenAI has published new system-prompt templates and API guidance for GPT-6.2 Astra and Sol, the models it shipped on November 17, and the message to developers is that the old instruction stacks are now part of the problem. The guidance, detailed by The Decoder, covers writing style, initiative, and coding, alongside Responses API changes that can silently discard reasoning content and drop two long-standing sampling parameters.
The prompt stack is now a drag
OpenAI says Astra writes more briefly and directly by default than earlier models, so instructions like "be concise" are redundant. Its recommended templates for everyday work ask the model to adopt terminology from the user and the source material, use active voice, and put key points first.
Two behaviors OpenAI observed in both models get dedicated countermeasure prompts. On long tasks, the model sometimes leaks shorthand from its own reasoning or from subagent messages into the final answer. The template instructs it to treat the final answer as the user's first look at the work, fully self-contained, and never a continuation of collapsed intermediate commentary.
Second, when intent is unclear, GPT-6.2 sometimes suggests a next step and waits. OpenAI now recommends telling the model to take authorized next steps itself, make reversible assumptions and flag them, and stop only for a real scope change, a concrete blocker, a user-owned decision, or an irreversible action. Higher reasoning effort helps the model follow through on final checks during long tasks, the company adds.
For coding, the template steers toward repo conservatism: reuse existing classes, methods, and APIs, avoid unnecessary helpers and abstractions, and skip new tests unless requested, required by the repo, or needed to catch a regression existing tests would miss. Fewer lines, shorter variable names, and minified formatting are explicitly called out as not being simplicity.
The API changes matter more than the prompts
Starting with GPT-6.2, the Responses API discards earlier reasoning content when a developer edits conversation history, switches to certain other models, or changes the end user or organization. The model then repeats work it had already done, which raises token costs and can degrade response quality. Integrations that thread reasoning across turns need to check for this.
Sampling parameters are also moving. Temperature and top_p are no longer supported on GPT-6.2. OpenAI recommends the medium reasoning level for new integrations and says existing setups pinned to none or minimal should migrate to low.
This is not OpenAI's first cleanup pass. In April the company advised against carrying GPT-5-era prompt stacks to GPT-5.5, in September it shipped prompts to curb GPT-6 Astra's follow-up question habit, and researcher Eric Provencher warned that bloated skills and AGENTS.md files slow the model. The through-line is that each generation needs fewer instructions, not more.
Why this is worth reading carefully
The guidance is an admission in operational form: the defaults changed enough between 6 and 6.2 that existing production prompts will behave differently, in ways that mostly will not raise errors. A pipeline that relied on "be concise" gets the same output with different prose. One that relied on the model stopping to ask before acting gets silent action. One that counted on reasoning continuity across edited histories gets recomputed tokens at full price. Independent personality-index measurements published this weekend show the same shift from the outside, with 40 percent shorter answers and near-disappeared hedging in Astra's output.
Teams running agentic workloads on GPT-6.x should diff their system prompts against the new templates before switching model IDs, budget for reasoning-content loss in any flow that rewrites history, and drop the sampling-parameter config that the API now ignores.
Sources
- The Decoder: OpenAI shares new prompting tips for GPT-6.2 Astra and Sol to fix brief answers and unnecessary questions
- OpenAI: Introducing GPT-6.2 Astra and Sol
Frequently asked questions
Why should developers shorten prompts for GPT-6.2?
OpenAI states that GPT-6.2 Astra writes more briefly and directly by default, making older instruction stacks like "be concise" redundant and a drag on performance.
What sampling parameters are no longer supported in GPT-6.2?
The temperature and top_p parameters are no longer supported on GPT-6.2; OpenAI recommends migrating existing setups to the low or medium reasoning levels.
How does the GPT-6.2 Responses API handle reasoning content?
Starting with GPT-6.2, the Responses API discards earlier reasoning content when a developer edits conversation history, switches models, or changes the user or organization, which can raise token costs.
