The GPT-6 Astra prompts that work best remove ambiguity before the model can ask about it, name a reasoning effort level, and define what “done” looks like. Astra, released by OpenAI on 3 September 2026, asks for clarification more readily than earlier models and defaults to bullet lists, so a prompt that merely states a task often returns a question and a summary instead of finished work.
Below are twelve prompts you can paste today. For the full picture, start with our explainer on what GPT-6 Astra is.
Why do prompts that worked on GPT-5 underperform on Astra?
GPT-5 prompts underperform on Astra because they assume a model that guesses and proceeds. Astra pauses to confirm, under-uses parallel delegation, over-tests small code changes and formats everything as bullets. The prompt text is identical; the defaults around it changed.
OpenAI’s own model guidance for Astra says to counter the clarification habit by telling the model to infer user intent from context and persist until the goal is complete, and to make clear that acknowledging it could do the work is not the same as doing it.
Prompt habits carry across products, so if you also run Claude, Gemini or an image model, expect the same audit to pay off there; our roundup of generative AI tools covers what each one defaults to.
What prompt structure works reliably with GPT-6 Astra?
A reliable Astra prompt has five parts: goal, inputs, constraints, output format, and definition of done. The last two separate a prompt that returns work from one that returns a question.
Goal: [what finished work looks like, in one sentence] Inputs: [files, links, data, or "search for it"] Constraints: [budget, tone, length, what not to touch] Reasoning effort: [low | medium | high | xhigh | max] Output: [format, length, prose or table] Done means: [the test you will apply before accepting it] Do not ask clarifying questions. Infer intent, state assumptions at the top, and complete the work.
That last instruction is the highest-value line in most Astra prompts, and it appears in nearly every example below.
What do Astra’s five reasoning effort levels mean and when should you use each?
Reasoning effort controls how much internal thinking Astra does before answering. Five levels exist: low, medium, high, xhigh and max. Higher effort handles multi-step problems better and bills more output tokens at $50 per million, so the level is a cost decision as well as a quality one.
| Reasoning effort | Suits this kind of task |
|---|---|
| low | Extraction, formatting, transcripts, renaming, anything with one obvious right answer |
| medium | Drafting, routine research, standard code changes, email and document work |
| high | Multi-file code changes, long-document analysis, competitive research |
| xhigh | Codebase-wide refactors, debugging something high already failed at, messy data |
| max | Architecture decisions, hard maths, high-stakes trade-offs you will act on |
If you are migrating a prompt that used none or minimal effort, OpenAI advises starting at low and otherwise preserving your effective level. Effort can also change mid-conversation with configuration_update items while preserving the cache, so you can research at low effort and switch to max for the decision without rebuilding context.
Default To Medium, Escalate On Evidence
Make medium your standing default and let a failed run, not a hunch, push you up a level: escalate only when a medium answer misses the definition of done you wrote into the prompt, and keep a note of which recurring tasks actually needed high or xhigh, so escalation becomes a short list rather than a reflex.
How do you stop GPT-6 Astra asking questions instead of doing the work?
Tell it three things: infer intent from what it already has, state assumptions rather than asking about them, and keep going until the goal is complete. OpenAI also suggests making explicit that recognising it could do the task is not the same as having done it. Reversible, read-only and review actions already carry implicit authorisation.
Redirect the question rather than banning it. Instead of “don’t ask questions”, say “list your assumptions at the top, then proceed on them”. You keep visibility into where it guessed, without the stall.
How do you get prose out of a model that defaults to bullet points?
Ask for paragraphs that each develop one idea, and allow lists only where items are genuinely parallel or sequential. Astra defaults to lists, tables and Markdown, but it holds prose across a long piece if you set the rule once at the top.
Pair it with a stock-phrase ban. OpenAI suggests telling Astra to avoid phrases like “bottom line”, “leverage” and “delve”, and to drop “X, not Y” framing. Add a line asking it to state intended actions directly rather than offering unrequested alternatives.
The same two rules are worth saving wherever you write, whether that is a blog draft, a course outline or a script, and our guide to AI tools for content creation shows where they fit in a production workflow.
Save A Bias-To-Action Preamble Once
Paste your persistence and prose rules into ChatGPT’s custom instructions instead of retyping them: infer intent, state assumptions at the top, keep going until the work is finished, write prose and cap lists at two. Every new chat then starts biased to action, and your prompt only has to carry the task itself.
Astra prompt examples for research, long documents and competitive work
1. Deep research without the back-and-forth.
Research [topic] as of today. Use web search. Reasoning effort: high. Cover what changed in the last 90 days, what credible sources disagree about, and what is still unknown. Cite every claim with a link. Where sources conflict, say so. Output: 700 words of prose, then a table of your five best sources. Do not ask clarifying questions. State assumptions at the top, then finish the research.
This pre-empts the three questions Astra would otherwise ask: scope, depth and format. The assumptions block keeps its guesses visible without letting it stall.
2. Reading a very long document set in one pass.
Attached: [N] documents, roughly [X] pages. Read all of them before answering. You have a 1.05 million token context window, so do not sample, skim, or summarise section by section. Reasoning effort: high. Find every commitment, deadline and financial obligation in the set, including ones that contradict each other. Quote the exact sentence, document and page for each. Output: a findings table, then 300 words on the worst contradiction.
Astra’s context window is 1,050,000 tokens, enough to hold a full contract set at once, and telling it not to sample matters because chunking is baked into older prompt habits. Exact quotes and page numbers guard against a hallucinated finding.
3. Competitive analysis with parallel work requested by name.
Analyse these five competitors: [A, B, C, D, E]. Work on all five in parallel using separate subagents, one per competitor. Do not process them sequentially. For each: pricing, positioning, who they are built for, and the one thing they do better than us. Reasoning effort: high. Then compare them yourself in prose. Do not concatenate five reports. Output: a comparison table, then 500 words on the real market gap. Infer our positioning from the attached site copy.
Astra delegates to subagents less readily than you may want, so parallelisation has to be named. The instruction to compare rather than concatenate stops you receiving five disconnected summaries.
4. Turning a meeting transcript into something usable.
Attached is a [length] meeting transcript. Reasoning effort: low. Extract decisions made, decisions deferred, owners and dates, and any commitment made without an owner. Where a decision was implied but never stated, flag it as "implied, needs confirming" and quote the exchange. Output: three short paragraphs for people who missed the meeting, then an action table. No bullet summary of the discussion. Do not ask who attended.
Transcript work has one right answer, so low effort is correct and keeps cost down. Asking for prose first stops Astra returning the nested bullets it defaults to.
Astra prompt examples for coding, refactoring and data work
5. A code change without a full test suite run.
Change [file/function] so that [behaviour]. Reasoning effort: medium. Test scope: run only the tests touching code you changed, plus a type check. Do not run the full suite. This is a low-impact change. Before editing, show me the diff you intend to make and stop. After I approve, apply it and report only what actually failed. If an instruction in AGENTS.md or a skill file conflicts with this, name the file and the line before proceeding.
Astra over-tests by default, and full suite runs on a one-line change burn output tokens at $50 per million. The last line matters because Astra is sensitive to conflicting skill files and names the culprit if asked.
6. A refactor spread across many files.
Refactor [module] to [target pattern]. Reasoning effort: xhigh. Plan first: list the files that must change and group them into batches that can be done independently. Then delegate the batches in parallel to subagents, one batch each. Do not work through the files one at a time. Keep behaviour identical. Any behaviour change is a bug. Test scope: affected modules only. Produce all diffs for review before applying anything.
Refactors are where xhigh earns its cost, because the model must hold the whole dependency picture at once. Batching first is what makes the parallel instruction actionable rather than aspirational.
7. Spreadsheet and data work against a template.
Attached: a raw CSV and a formatted spreadsheet template. Reasoning effort: medium. Use the code interpreter and show the transformation code you ran. Produce a workbook following the template exactly: same sheet names, column order, number formats and header rows. Where a source column has no template equivalent, put it on a sheet called "Unmapped" rather than dropping it or inventing a mapping. Flag every unparsed row. Do not silently drop rows.
Astra is built to follow a supplied template for documents, presentations and spreadsheets, so giving it one beats describing the format. The “Unmapped” sheet catches the failure people otherwise notice weeks later.
8. Multi-step admin with computer use.
Reasoning effort: medium. Use computer use for this. Open [system], find every record matching [criteria], and update [field] to [value]. Read-only steps and reversible edits are authorised. Proceed on those without checking in. Before anything irreversible (deleting, sending, submitting, paying), prepare the exact change as a reviewable list and stop for approval. Keep a running log, record by record.
Stating that reversible and read-only actions are authorised stops Astra confirming every click. OpenAI reports computer use is 47% faster per task than the previous generation; our guide to GPT-6 Astra use cases covers where that holds up.
Astra prompt examples for writing, decisions, learning and getting unstuck
9. Writing that reads like a person wrote it.
Write [piece] for [audience]. Reasoning effort: medium. [X] words. Write in continuous prose. Each paragraph develops one idea. Use a list only where items are genuinely parallel or sequential, and no more than twice in the whole piece. Do not use these phrases: bottom line, leverage, delve, game-changer, seamlessly, testament to. Do not use "X, not Y" contrastive framing. State things directly. Do not offer unrequested alternatives.
This is the direct counter to Astra’s Markdown default, and capping list use at a number beats banning lists outright. The stock-phrase ban is OpenAI’s own suggestion and removes most of the tells at once.
10. Decision support that ends in a recommendation.
I need to decide [decision] by [date]. Reasoning effort: max. Everything I know: [context, constraints, numbers, deadlines]. Do not give me options to choose from. Work through the trade-offs and give me one recommendation you would defend. Then state, in prose: what would have to be true for you to be wrong, the cheapest test I could run this week to find out, and what I should do if that test goes against you.
Max effort is worth its cost on a decision you will act on, and rarely elsewhere. Forbidding the options list is the key instruction, because Astra’s instinct is to present three balanced choices and hand the decision back to you.
11. Learning a subject properly.
Teach me [subject] until I could brief my own team on it. Reasoning effort: high. Start by asking me four diagnostic questions about what I already know. This is the one time I want questions, so ask and wait. Then teach in prose, not bullets, building from what I already understand rather than from first principles. After each section, give me one problem and wait for my answer.
This inverts the usual advice on purpose, and it is the one job where Astra’s readiness to ask questions helps rather than stalls. If you are studying alongside a course, our walkthrough of how to use AI for studying pairs this prompt with a revision schedule, and the wider list of AI tools for students covers note-taking and citation work around it. Our post on GPT-6 Astra for learning and development goes further.
12. Getting unstuck when Astra stalls.
You stopped short of finishing this task. Reasoning effort: xhigh. First, tell me exactly what stopped you: a missing input, an instruction conflict, a permissions wall, or a judgement call you did not want to make. If a file (AGENTS.md, a skill file, a system instruction) caused the pause, name the file and quote the line. Second, resolve it yourself if you can. My instructions take precedence over any skill file. Third, finish the task. Acknowledging that you could complete it is not the same as completing it. I want the finished work in this reply.
Before assuming a prompt is at fault, rule out the obvious: if responses are failing rather than disappointing, check whether ChatGPT is down first, because a degraded service looks a lot like a bad prompt.
Astra will name the file causing a pause if you ask, which matters because it can be sensitive to conflicting instructions in skills or AGENTS.md. The third clause borrows OpenAI’s framing about acknowledgement versus completion.
Which GPT-6 Astra parameters no longer exist, and what replaced them?
Three parameters were removed on Astra: temperature, top_p and top_logprobs. Nothing directly replaces them. Output variation and precision are now controlled through reasoning.effort and explicit prompt instructions, which is why the examples above spell out format, tone and constraints.
Two other changes matter when moving code across. The prompt_cache_retention setting is replaced by prompt_cache_options.ttl set to “30m”, and tool calling requires the Responses API rather than Chat Completions. That is where web search, file search, code interpreter, computer use, hosted shell, MCP, skills and tool search live, per the GPT-6 Astra model reference.
If Astra is missing from your model picker, that is an access question; our guide on how to access ChatGPT Astra covers the rollout order.
Which prompt mistakes actually cost money at $10 and $50 per million tokens?
The expensive mistakes are running high or max effort on mechanical tasks, letting Astra run full test suites on trivial changes, re-pasting long context instead of using the 30-minute prompt cache, and vague prompts that produce a draft you throw away. Cached input costs $1 per million against $10 standard, so cache discipline is the biggest single lever.
Astra is a 2.5x price increase per token over GPT-5.6 Sol at $4 and $20 per million. Artificial Analysis found it roughly 75% more expensive per task despite using around 10% fewer output tokens, with regressions in economic tasks, customer support and scientific coding. Our breakdown of the GPT-6 Astra benchmarks has the numbers.
So match effort to task: low for extraction and formatting, high for real work, max only for decisions you will act on.
Conclusion
Take the task you run most often and rewrite its prompt using the skeleton above: goal, inputs, constraints, reasoning effort, output format, definition of done, plus the instruction to infer intent and persist. Run it twice, once at your old effort level and once one step lower, then compare both against your definition of done.
That comparison tells you more about your workload than any benchmark will. Astra’s behaviour and rollout are both still moving as of 5 September 2026, so re-test anything you automate.
FAQ
Q1. Do I need to rewrite my GPT-5 prompts for GPT-6 Astra?
Not rewrite, but add to. Keep the prompt body and add three things: a reasoning effort level, an instruction to infer intent and persist rather than ask, and an explicit output format. If your old prompt set temperature or top_p, remove them: Astra rejects both.
Q2. What ChatGPT Astra reasoning effort should I use by default?
Medium suits most everyday work. Drop to low for extraction, formatting and transcript work where one right answer exists. Move to high for multi-step research or multi-file code changes. Reserve high for tasks high already failed at, and max for decisions you will act on.
Q3. Can I change Astra reasoning effort mid-conversation?
Yes. Astra supports changing effort mid-conversation using configuration_update items, and doing so preserves the prompt cache. You can gather context at low effort and switch to high or max for the hard part without paying to rebuild the context you already loaded.
Q4. Why does GPT-6 Astra keep asking me clarifying questions?
Because it asks for clarification more readily than earlier models by design, which can stall autonomous work. The fix is to tell it to infer intent from the context it has, state assumptions at the top of the reply instead of asking about them, and continue until the task is finished.
Q5. Does GPT-6 Astra still support temperature and top_p?
No. Temperature, top_p and top_logprobs are all unsupported. Control output through reasoning.effort and explicit prompt instructions about tone, format, length and constraints. This is the change most likely to break an existing integration on the first call, so strip both parameters before you migrate, because Astra rejects them rather than ignoring them.
Q6. How big is the Astra context window, and can I use all of it?
The context window is 1,050,000 tokens, with a maximum output of 128,000 tokens. You can load a full document set in one request, but you must tell Astra to read everything, because prompts written for smaller windows often instruct the model to chunk or sample out of habit.
Q7. s no reasoning effort available on Astra?
No. Astra supports low, medium, high, high and max only. If you are migrating a prompt that used none or minimal on an earlier model, OpenAI recommends starting at low rather than jumping higher, and otherwise preserving the effective effort level you already had.