GPT-6 Astra: what changed and when to pay for it

On this page

1. 1.What GPT-6 Astra is, and where you can get it
2. 2.What changed, and what did not
3. 3.What it costs
4. 4.When to use it, and when not to
5. 5.How to brief it: prompts get shorter, not longer
6. 6.What the safety report means for your governance
7. 7.Six claims you will hear this week
8. 8.What to do in the next 30 days
9. 9.Where this article stops
10. 10.Where to go from here

OpenAI released GPT-6 Astra on 3 September 2026. If you run an SME, a charity or a public sector team, the short answer is this: do not move all your work to it. Move the work where failure, rework or long multi-step tasks cost more than the higher price.

This guide is for leaders deciding whether to pay for it, and for the people who will write the prompts. We wrote it one day after launch, so nearly all the performance evidence comes from OpenAI itself. We say where that matters.

We do not sell OpenAI licences, or anyone else’s. The 30-day test at the end works for any vendor’s model.

What GPT-6 Astra is, and where you can get it

GPT-6 Astra is OpenAI’s new top model for hard, multi-step work: reasoning, coding, operating a computer, research and producing finished documents. OpenAI has not published its size or how it is built. Nobody outside OpenAI can tell you either.

You will not see the name Astra everywhere. In the API, the developer route for building software on the model, it is called gpt-6-astra. In ordinary ChatGPT it appears as GPT-6 Pro on eligible Pro, Business and Enterprise plans while the rollout continues.

Two access points need checking before you promise it to anyone:

• OpenAI’s launch article said Astra would reach Plus users in the days after launch. OpenAI’s current help article says GPT-6 Pro in ordinary Chat is not included with Plus. Trust the help page and your own workspace settings, not the launch post.
• Enterprise access was switched off by default at launch. An administrator has to enable it.

One rule saves a lot of confusion. What the model can do in the API is not what it can do in a chat window. Tools, permissions and controls differ by product, so check the product you use rather than the one in the announcement.

What changed, and what did not

The context window did not change. A context window is the amount of text a model can work with in one go. It is measured in tokens, and a token is roughly three quarters of a word.

Astra and the three GPT-5.6 models it sits above, Sol, Terra and Luna, all document a 1,050,000-token window and a 128,000-token maximum output. What OpenAI says improved is how the model uses that space. It picks out the relevant material, stays on task across long jobs, operates software and produces finished documents.

That is a claim about behaviour, not capacity, so “a bigger window” is the wrong way to describe this release.

OpenAI’s launch figures show big gains in some areas and small ones in others. The table shows a selection, comparing Astra with Sol, the GPT-5.6 flagship. These are OpenAI’s numbers from OpenAI’s test setup, and no one has independently reproduced them yet.

Read the bottom three rows before the top four. On browsing, software engineering and the broad intelligence index, the gap is one or two points. The strong evidence is for computer operation, hard coding, professional artefacts and very long documents.

What did not change matters as much:

• It still makes things up. OpenAI reports fewer errors, not none.
• Its knowledge stops at 30 April 2026. Anything later needs web access or documents you supply.
• Prompt injection, where text inside a document or web page hijacks the model’s instructions, is improved and not solved. More on that below.
• The person who signs off the work is still accountable for it.

What it costs

Astra is priced well above the models it sits alongside. The table shows OpenAI’s standard API prices per million tokens at launch.

That makes Astra 2.5 times Sol’s price on every line, 5 times Terra on input and 50 times Luna on input. Sol’s price is a promotion, so recheck it before budgeting.

OpenAI’s counter-argument is that Astra can finish a task in fewer tokens, so the cost per completed task can be lower. Its launch material gives examples: about 43% cheaper on BenchCAD and about 9% cheaper on Terminal-Bench 4.0 in the setups compared. Those are chosen examples, not a rule.

Our arithmetic makes the trade-off concrete. Take a task with the same mix of input and output on each model. Astra has to finish in about 40% of the tokens Sol would use to break even on token cost alone.

Where it cannot, it has to earn the difference through fewer retries, less human editing or a better outcome.

One more pricing rule catches people out. Above 272,000 input tokens, higher multipliers apply to the whole request. A big window is not an invitation to attach everything you own.

When to use it, and when not to

Our rule for clients is one sentence. Do not route work to Astra because it is GPT-6. Route work to Astra when failure, retries, complex tool use, hard judgement, long-running tasks or human rework cost more than the extra tokens.

The table is our starting recommendation by type of work. It is our analysis, built on OpenAI’s own positioning of each model. OpenAI places Terra at the balance of capability and cost, Luna on high-volume work and Astra on the hardest end-to-end tasks.

Three types of work should not go to Astra at all. Use ordinary software where the rule can be written down exactly: arithmetic, validation, access control and transaction rules. Keep a person in charge where accountability is the point: employment decisions, legal judgement, spending authority, health and safety.

Keep Sol where you cannot show that Astra’s extra success rate pays for a 2.5 times price.

How to brief it: prompts get shorter, not longer

OpenAI’s developer guide for Astra lists behaviours that should change how you write prompts. The points that matter most for office and consultancy work are these:

• It asks more questions when missing information could change the result. Left alone, it may stop where an older model would have guessed.
• It follows instructions more closely, including instructions in files like AGENTS.md and reusable Skills. Bad instructions work better too.
• It writes long, heavily formatted Markdown by default, with recurring phrases.
• It delegates to sub-agents less than you might want unless told to.
• It can over-test small code changes.
• It holds an objective across long tasks and copes with a change of requirements part way through.

The consequence is not bigger prompts. A good Astra prompt says five things:

• what the result is;
• which sources win when they disagree;
• what the model may assume on its own, and when it has to ask;
• what it has to check before finishing;
• what “done” looks like.

Ours look like this:

Outcome: what has to be accomplished.

Use the material provided. Infer routine, reversible details rather than
asking. Ask only when missing information could materially change the result.
Treat the signed policy as authoritative. Use meeting notes as background only.

Deliver: the output and its format.
Done when: an observable completion condition is met.

Three habits from the GPT-5 era should go. A long invented persona, “you are a senior strategist with 30 years’ experience”, adds nothing the task does not already state. “Think step by step” has no Astra-specific evidence behind it, and the model has its own reasoning-effort setting.

Repeating an instruction three times used to help. With a model that follows instructions closely, it now creates contradictions.

One clause is worth adding to almost every client-facing prompt: “Write in plain professional prose. Use a table only for a genuine comparison. Do not add generic introductions, repeated summaries or canned transitions.” That is a direct response to the formatting habit OpenAI describes.

What the safety report means for your governance

OpenAI’s System Card for Astra reports two things at once. Read them together.

Astra is the first OpenAI model to meet the Critical cybersecurity threshold under OpenAI’s own Preparedness Framework. OpenAI defines that as the capability to find unknown flaws and build attack paths across many hardened systems. It assumes the model has tools and access, and no person directing each step.

That is OpenAI’s internal classification, not a legal one. OpenAI says it found previously unknown vulnerabilities during testing, and has added safeguards, refusals and monitoring in response. Those safeguards can pause or stop a legitimate task, so build that into your process.

At the same time OpenAI reports stronger alignment than Sol: it respects restrictions better and resists prompt injection better. It also reports a harder-to-monitor model. Under adversarial testing, Astra is better at controlling what shows up in its own reasoning trace.

OpenAI says this came mostly from adversarial settings and that Astra breaks restrictions less often overall.

Our reading is short. Better alignment lowers the chance of a problem, and greater capability raises the cost of one. Governance still matters, and the argument “it is safer, so we can drop human approval” fails.

Prompt injection deserves its own line. A model that ignores malicious instructions 99 times in 100 still fails once in 100, and an attacker gets to try repeatedly. So keep the controls you already have:

• treat text from documents and web pages as data, not instructions;
• give the model the fewest tools it needs;
• keep write permissions narrow;
• log what it does;
• check the result after the fact.

For any agent, we recommend a person approves before the model does any of the following:

• moves money;
• accepts legal terms;
• sends a sensitive external communication;
• deletes or overwrites important data;
• changes who has access to what;
• deploys to production;
• disables a security control;
• makes a regulated decision.

Everything reversible below that line can run on its own once you have tested it.

Six claims you will hear this week

What to do in the next 30 days

Do not change your default model this month. Run a test instead. This is the method we use before recommending any model change, whichever vendor makes it.

1. Pick 30 to 50 real tasks from your own recent work, with personal data removed and a clear pass mark for each.
2. Run each task on your current model and on Astra with the same prompt. That tells you whether the gain comes from the model or from a rewritten prompt.
3. Record five things per run: pass or fail, minutes of human editing, tokens used and the number of clarifying questions. The fifth is any approval point the model should have stopped at and did not.
4. Work out cost per accepted deliverable, not cost per token. Include retries, review time and correction time.
5. Audit every file that carries instructions into the model: AGENTS.md files, Skills and project instructions. Check who owns each, whether it is current and whether two of them contradict each other.
6. Recheck pricing, the Sol promotion and Plus availability at the end of the month. All three were unsettled at launch.

Only then decide whether any type of work moves to Astra by default.

Where this article stops

Everything above rests on OpenAI’s model documentation, developer guide, help pages and System Card as published on 3 and 4 September 2026. Independent testing did not exist when we wrote it. Treat the benchmark figures as a guide to where to test first, not as a promise of what you will get.

We cannot tell you Astra’s size or its general speed advantage. We cannot tell you its real-world prompt-injection failure rate, or how long it stays reliable on a task that runs for hours. Nobody outside OpenAI can yet.

Put three questions to your own supplier or IT lead before you commit:

• Which plan are we on, and is GPT-6 Pro included?
• Has our workspace administrator enabled it?
• Does our contract cover the API’s zero data retention option if we handle personal data?

Where to go from here

If you use ChatGPT today, our guide to ChatGPT for business covers the governance basics that do not change with the model. If you are planning agents that act rather than advise, read AI agents: the next evolution in knowledge work first.

If you want the 30-day test run on your own tasks, our AI at Work engagements do exactly that. Each workflow ends with a governance pack. Book a free discovery call.

Back to top