Guides

How to Prompt GPT-6.2 Astra and Sol: What to Change

What OpenAI's GPT-6.2 guide says about style, autonomy and caching, with our own before-and-after prompts you can copy for Astra and Sol.

GPT-6.2 Astra and GPT-6.2 Sol arrived in OpenAI’s API on November 17, 2026. OpenAI’s migration and prompting guide says the models behave differently from earlier ones in a few ways that affect how you write prompts. The facts below come from that guide and OpenAI’s model pages. The prompts and the advice on applying them are our own.

What OpenAI says changed

  • Shorter by default. OpenAI says GPT-6.2 is more concise and direct, so it suggests revisiting “be concise” instructions.
  • It may stop and wait. Without guidance, the model may propose a next step and wait for confirmation. For unsupervised work OpenAI suggests stating the outcome, asking it to keep going within scope and to flag assumptions.
  • Reasoning effort: start at medium; if you used none or minimal before, use low (none is rejected).
  • Instructions placement: top-level instructions are now appended after the input. Put reusable instructions in a leading developer message so they can be cached.
  • No more sampling knobs: OpenAI says to remove temperature and top_p when migrating, so tone and variety now have to come from the prompt itself.

Which model to use

JobModelWhy
Hard reasoning, big refactors, long agent runsGPT-6.2 AstraOpenAI’s pick for the most demanding tasks; $8 / $40 per 1M tokens
Everyday writing, coding, agents at scaleGPT-6.2 SolOpenAI’s go-to for complex work at lower cost; $2 / $10
Overnight or bulk jobsEither, with a 2h or 8h windowDiscounted background windows cut token rates by 30% or 40%

Our rule of thumb: run your real prompt on Sol first. Move to Astra only for the cases where Sol’s answer fails your own check.

Prompt 1: an outcome-first task brief

Before (older style):

Be concise. Look at this landing page copy and suggest improvements.

After (written for GPT-6.2):

Goal: rewrite the landing page copy below so a first-time visitor understands
the product in one read.
Audience: small online shop owners who have never used AI image tools.
Keep: the product name, the price and the three feature names exactly as written.
Work through it without stopping for my approval. If something is unclear,
make a reasonable assumption, list it at the end and carry on.
Stop and ask only if the copy contradicts itself on price or features.
Before you finish, check that every section of the original has a rewrite.
Output: the new copy, then a short list of assumptions.

The brief does what OpenAI recommends: it names the outcome, gives permission to keep going within scope, says when to stop, and asks for a final self-check.

Prompt 2: a reusable developer message

Put stable rules in one developer message at the start of the input, and keep request-specific details out of it, so the prefix stays the same between calls.

You are the writing assistant for a small design studio.
Use the client's own terms for their products. Use active voice.
Lead with the key point. When a draft is for an outside reader, add the
background they would not know. Final answers must make sense on their own,
without referring to your earlier notes or progress updates.

Prompt 3: adding back variety without temperature

With sampling settings gone, ask for the spread directly:

Give me five headline options for this product. Make them clearly different:
one plain and factual, one playful, one question, one built around a number,
one under five words. Label each style.

Prompt 4: a coding request with your conventions

Fix the bug where the newsletter form submits twice.
Our conventions: reuse the existing form helper, no new dependencies,
keep functions short and readable, no extra abstraction layers.
Callers rely on the onSubmit event firing exactly once, so keep that behaviour.
Add a test only for this regression. Explain the cause in two sentences.

OpenAI’s guide suggests describing your conventions and the behaviour existing callers rely on, and adding tests only when requested or needed for a regression.

Quick checklist

  1. Delete old “be brief” lines unless you need very short answers.
  2. Say what done looks like, and when the model should stop and ask.
  3. Move stable instructions into a leading developer message.
  4. Do not edit earlier turns of a conversation you want to continue: OpenAI says this can drop reasoning and cost more tokens.
  5. Start reasoning effort at medium and only raise it when a task fails your check.

For agents that act on your behalf, add the limits from our AI agent permissions and approvals checklist. For more reusable structures, see AI prompt templates you can reuse and the AI prompt workflow.

stay in the loop

New guides, prompts and AI model news by email