How to prompt GPT-6.1 Sol

GPT-6.1 Sol is the newest modelModelA specific version of an AI, such as Claude Opus 5.5 or GPT-6 Astra. Different models trade speed, cost and ability. in OpenAI’s GPT-6 family. OpenAI released it on 29 September 2026 and places it close to GPT-6 Astra on complex professional work, at about one-fifth of Astra’s price. For most lecturers, researchers and small teams, that makes it the sensible model to try first.

One honest note. OpenAI has not published a separate prompting guide for Sol. Its main GPT-6 guide was written from watching Astra, and it asks you to test the advice on your own model and tasks. So this guide marks what is documented and what is our own suggestion.

✦

1

The idea in 30 seconds

Try it first

Near-Astra quality at about a fifth of the price.

Watch the 30-second explainer

The idea in 30 seconds0:00 / 0:31

1. Try it first Near-Astra quality at about a fifth of the price.
2. Say what "done" looks like The finished result, its format and how to check it.
3. Say when to ask Or it may stop halfway to check in with you.
4. Match the thinking level A quick glance for small jobs, a full audit only for hard ones.
5. Still verify Check figures and references, and remove personal data first.
6. Same habits as Astra, at a fraction of the cost. Move up only if the result falls short.

Key takeaways

If you remember nothing else, remember these.

• Sol is the default to try first. Move up to GPT-6 Astra only when Sol’s result on your task is not good enough.
• The core habits match Astra’s: a clear goal, a defined “done”, a check step and rules for when to ask.
• Pick the thinking level to fit the job. A higher level is slower and does not fix a vague request.
• In our view, prompts tuned tightly for a cheaper model may over-constrain Astra, so loosen them if you switch up.
• Check every result, and remove personal data before you paste anything.

Is this model for me?

Most people meet GPT-6 through ChatGPT, not by choosing a model by name. If you are new to writing prompts, start with the AIM method: a clear request matters more than which model answers. Come back here when you want to get more out of the GPT-6 family at a sensible cost.

Here is how Sol sits in the current lineup, in plain terms.

Our recommendation:

• Sol for most real work: drafting course materials, analysing a set of papers, writing reports and memos.
• Astra when Sol’s output on your own task is clearly not good enough, or for long, high-stakes work. OpenAI suggests comparing the two on your own tasks.
• Luna for simple, repetitive jobs at volume, such as sorting anonymised feedback comments into themes.

Why start with the cheaper model? If your department pays for API use, or you run the same task every week, a model at a fifth of the price adds up quickly. OpenAI describes Sol as close to Astra on complex work, so with a clear request the difference may be small on everyday tasks. Test it on your own work rather than assuming.

Sol’s knowledge cutoffKnowledge cutoffThe date after which the AI learned nothing new. is 30 April 2026. For anything newer, paste the official text or let it search. The model hub compares every current model from OpenAI and Anthropic.

Anatomy of a good Sol request

Here is a complete request that uses every habit in this guide. Tap each coloured part to see what it does.

Dissect a prompt

A Sol request for an accounting lecturer

ExploreTest yourself

Tap a coloured part of the prompt to see what it does.

You are an accounting lecturer at a Malaysian polytechnic who writes clear practice questions. My Week 6 notes on the statement of cash flows are inside <notes> tags below. Write eight practice questions on classifying cash flows. Done means eight questions, each with a worked answer in RM, covering operating, investing and financing items. Before you finish, recheck every figure so each worked answer adds up. Do not ask me questions first; note any assumptions at the end. Plain text, no headings.

The order follows OpenAI’s general advice: who the AI is, then the task, then examples, then the material. That is AIM in another form. Two or three varied examplesExamples (few-shot)Showing the AI one to five samples of the result you want, so it can copy the style and format. of what you want often teach more than a paragraph of description.

In ChatGPT: the habits that work

Sol in the API and the models inside ChatGPT are not necessarily the same thing. At the time of writing, OpenAI’s announcements suggest that ChatGPT’s paid plans run on GPT-6 Sol, the earlier Sol, and Free and Go plans on GPT-6 Luna (not confirmed on an official OpenAI page). Plans change often, so check the model picker in your own account. The habits below carry over whichever model answers.

Give the goal, not every step

OpenAI compares a reasoning model to a senior co-worker. You would not give a senior colleague a 12-step recipe for writing a memo. You would say what you need, for whom, in what form, and how you will judge it. A precise action verb states the goal sharply. If you are not yet sure what you need, the Chris Voss method helps you work it out in a short dialogue first.

Write a done contract

GPT-6 models may stop to ask clarifying questions or check in before the job is finished. A “done contract” prevents that: a few lines that say what the finished result is, its format, how to check it, and when to stop or ask.

For example, an office manager asking for a monthly report might add: “Done means one page, under 400 words, with three findings, each backed by a figure from the data. Recalculate every figure you quote. If something is missing, assume sensibly, note it at the end and keep going.” That is enough to stop most half-finished answers.

Pick a task below and tap each part of the contract. Watch which problems each part removes.

Interactive

Tell the AI what “done” looks like

Pick a task. Then tap each part below to add it to the prompt, and watch the risk list turn green.

Lecturer: rubricResearch: screeningOffice: travel claimsSME: sales report

Parts of a “done” brief

+DeliverablePrevents: stops too early+Format and stylePrevents: too much formatting+Verification stepPrevents: nothing checked+Stop and ask rulesPrevents: too many questions+Thinking levelPrevents: wasted time or cost

Tap a part to add it. Tap again to remove it.

Your prompt

Copy prompt

I teach Principles of Marketing (Diploma Year 1). CLO2: Present a marketing-mix (4P) proposal for a local SME product. Our faculty scale is Excellent (4), Good (3), Satisfactory (2), Weak (1).

Create a marking rubric for the CLO2 group presentation.

0

Risks covered: 0 of 5

Start with any part. The base request on its own leaves all five risks open.

• Stops too early
• Too many questions
• Too much formatting
• Nothing checked
• Wasted time or cost

Before: vagueAfter: precise

You are a qualitative researcher experienced in thematic analysis.

Twelve anonymised transcripts are inside <transcript id="P01"> to <transcript id="P12"> tags. The report is for a state SME agency and will be read by non-specialists.

Identify the main barriers to adopting AI tools. Done means four to six themes, each with a one-sentence definition, the supporting participant IDs and one short quotation, in plain paragraphs. Check every quotation against the transcripts before you finish. Do not ask questions first; list any assumptions at the end.

What you typically get back

Themes you can lift into your findings chapter, each traceable to participants, with quotations you can verify in seconds.

Ask for plain paragraphs

GPT-6 models lean towards MarkdownMarkdownSimple symbols (#, *, -) the AI uses for headings and bullets.: headings, bullet lists and tables. If the answer is going into an email or a report, ask for short, plain paragraphs, lists only when items are truly parallel, and no closing summary.

Three small habits that save time

• Name your attachments. “Use only the file Claims_Q3.xlsx” stops it mixing sources.
• Use a fresh chat for each new task, so earlier instructions do not leak in.
• Review old saved instructions. GPT-6 models may read rules such as “always ask me first” literally. Rewrite them to say when to ask.

Thinking levels: a quick glance or a full audit

Some ChatGPT plans let you choose how long the model “thinks” before answering. Developers call this reasoning effortEffort / reasoning effortA setting that tells the AI how hard to think before answering. Higher effort is slower and costs more, but helps with difficult work.. Think of an auditor: a quick glance at a receipt takes seconds, while a full audit of the year’s accounts takes days and costs far more. Each suits a different job.

These are our suggested starting points, not OpenAI rules. Test them on your own work.

Level names depend on your plan. OpenAI’s announcements appear to list options such as Instant, Medium, High and Extra High (not confirmed on an official OpenAI page). OpenAI’s guidance treats this setting as a tuning knob, not a fix for a weak request. It is tempting to turn it up on a cheaper model; fix the request first.

Ready-to-copy prompts for local work

Replace the Input details with your own, and remove personal data first.

You are an assessment designer for a Malaysian diploma programme who writes clear multiple-choice questions.

My lecture notes for Week 6 are inside <notes> tags. CLO2: Explain the factors that affect consumer buying behaviour. Students are Diploma Year 1.

Write 10 multiple-choice questions on CLO2. Done means 10 questions, each with four options, one correct answer, a one-sentence explanation and its revised Bloom's level (Remember, Understand or Apply), with at least four at Apply. Use only content from my notes. Before finishing, check that no two questions test the same point. Plain text, no table.

You are an experienced finance officer who checks staff claims against policy.

Three months of anonymised travel claims are inside <claims> tags, one row per claim with an ID. Our claims policy is inside <policy> tags.

Check every claim against the policy. Done means a list of only the claims that break a rule, each with the claim ID, the rule broken and the amount in RM, followed by the total value of flagged claims per month. Recalculate the monthly totals before you finish. If a claim is missing information, list it separately as "cannot check". Plain text, one short list, no headings.

You are a customer-service lead for a Malaysian online furniture shop.

Thirty anonymised customer messages, in English, Malay or a mix of both, are inside <messages> tags, each with an ID.

Sort every message. Done means a table with ID, category (delivery, damage, refund, product question, other), urgency (high, medium, low) and a one-line suggested reply in the customer's language. Count the rows before finishing; there must be 30. Mark urgency high for any message that mentions injury or a safety risk.

Moving a prompt up to Astra

How do you know Sol is not enough? Improve the request first, then look for these signs on your own task:

• It stays shallow. The analysis lists points but never weighs them, even with a clear done contract.
• It misses items. It skips rows or documents despite a count check.
• The stakes are high. A mistake would be costly, as in an audit finding or a funding report, and a second opinion is worth the price.

If you see these, try the same task on Astra and compare. Do not just paste the same tightly scripted prompt, though.

OpenAI’s Astra blog warns that guidance written for its other models (it names the older GPT-5.6 Sol) can overconstrain Astra. The blog does not mention GPT-6.1 Sol, which came out later. Still, we expect the same when you move a tightly scripted 6.1 Sol prompt up to Astra. So loosen it: keep the goal, the limits, the format and the done contract, and drop the step-by-step recipe.

For developers: technical detailsOptional: skip this if you only use the chat app

What OpenAI documents for Sol specifically (model page and “Using GPT-6”, checked 12 October 2026):

• Model ID: gpt-6.1-sol, the only snapshot listed today. Pin it in production and keep evaluations so you notice changes.
• Do not confuse it with gpt-6-sol, the earlier GPT-6 Sol released on 22 September. The two have different settings; only the newer one is covered here.
• Reasoning effort: low, medium (default), high, xhigh or max. Requests with none or minimal are not supported. If an older workflow used minimal, OpenAI suggests starting with low and comparing results on representative tasks.
• Price: US$2 input and US$10 output per million tokens, one-fifth of Astra’s US$10 / US$50. Cached input costs US$0.10, 5% of the normal input rate, so long, reusable instructions at the start become very cheap. Prompts over 272,000 input tokens are billed at higher rates for the whole request.
• Specs: context window about 1.05 million tokens, up to 128,000 output tokens, text and image input, text output.

Settings shared with the family:

• Layers. Rules go in the instructions parameter or a developer message, which outranks the user message. Our Astra guide has a diagram of the layers.
• Use the Responses API for tools. Chat Completions works without tool calling; tool calling needs Responses. At launch, multi-agent support for Sol was in beta.
• Remove sampling settings. With reasoning on, remove temperature, top_p and top_logprobs (and logprobs in Chat Completions).
• Change effort mid-conversation with a configuration_update input item, raising it for a hard step and lowering it for routine follow-ups. Keep the request-level effort unchanged so your prompt prefix stays cacheable.
• Instruction files. If skills or AGENTS.md files are shared by agents running on different models, think about which models will read them.
• Test before you switch. Run the same evaluation set on Sol and Astra and decide with evidence.

{
  "model": "gpt-6.1-sol",
  "reasoning": { "effort": "medium" },
  "instructions": "Identity, rules, done criteria and examples.",
  "input": "Today's request, then reference material in tags."
}

Check your understanding

Check yourself

Prompting GPT-6.1 Sol in five questions

1/5

Pick an answer, then tap Check answer. You can also use the arrow keys to move between answers.

Question 1 of 5: Sol returned half a rubric and asked whether to continue. Which part of the request was missing?

AA definition of done and permission to proceedBA longer description of who the AI should beCA higher thinking level

Check answer

Sources

8

1. 1Using GPT-6 (OpenAI API docs) (opens in a new tab)developers.openai.com
2. 2GPT-6.1 Sol model page (opens in a new tab)developers.openai.com
3. 3GPT-6 Astra model page (opens in a new tab)developers.openai.com
4. 4GPT-6 Luna model page (opens in a new tab)developers.openai.com
5. 5Rethinking skills and prompts for GPT-6 Astra (OpenAI developer blog) (opens in a new tab)developers.openai.com
6. 6Prompt engineering guide (opens in a new tab)developers.openai.com
7. 7Reasoning models guide (opens in a new tab)developers.openai.com
8. 8OpenAI API changelog (opens in a new tab)developers.openai.com

Nice work!

Guide complete

You made it to the end. Take the quiz above if you have not yet, for bonus XP.

• 0 XP
• 0 day streak