GPT-6 Astra is OpenAI’s newest frontier model, built less like a chatbot and more like a computer operator that uses software, browses the web, and carries long multi-step jobs through to a finished document, spreadsheet or presentation. OpenAI released it on September 3, 2026, priced at 2.5 times its predecessor GPT-5.6 Sol on the published per-token rates, and a single conversation can carry a little over a million tokens.
The coverage since launch has been benchmark scores, a debate about whether this counts as artificial general intelligence, and a cybersecurity story that mostly concerns security researchers. For anyone running a small business rather than a research lab, there are two questions: what work can you actually hand this thing, and what does the higher price actually mean once you’re using it?
Key takeaways:
- GPT-6 Astra costs 2.5 times what GPT-5.6 Sol did, which makes this the first frontier launch in a while where price is the decision rather than a footnote.
- It’s built to operate software rather than only produce text. It can do things like using applications, browsing, inspecting what’s on screen, building sites, running quality checks and finishing work that runs through several stages, which changes the question from which model writes better to what work you’d hand over entirely.
- OpenAI’s president has said he personally believes the company has reached artificial general intelligence, while independent researchers are not so convinced. Still, this doesn’t change what your business should do this month.
- The way to find out whether it earns the extra cost is to run one job you already do on Astra and on the cheaper model you use now, then compare the finished result.
What is GPT-6 Astra, and what makes it different?
Every model release comes with a claim about being smarter than the last one, and that claim is hard to act on. What’s different about GPT-6 Astra is the kind of thing OpenAI says it’s for.
It’s built to operate software, not just write text
OpenAI positions Astra as state-of-the-art on computer use, which in plain terms means it can work the way a person works at a screen: look at what’s on the display, decide what to do next, click or type, check what happened, and carry on. One thing to be clear about before you get excited: the model decides the actions, and the application around it supplies the access and carries them out.
Picking Astra from a dropdown doesn’t hand it your desktop. OpenAI’s own examples are filling out online forms, updating customer records in a CRM, organizing your calendar, researching something online and drafting the summary straight into your email or document editor, analyzing data and generating the plots, building a website and then running quality checks to confirm its features work, and installing and testing software while troubleshooting what it sees on screen.
For a small business actively using AI, that’s a different proposition from a better writer. A writing model gives you a draft you then do something with. A model that operates software can take a job that has several stages, do each stage, and hand you the finished thing at the end. GPT-5.6 Sol could already do this sort of thing, and OpenAI scores it on the same computer-use tests, so the claim for Astra is that it does the work better rather than that it does work Sol couldn’t. OpenAI reports it finishing computer-use tasks in roughly 47% less time, at around 40 minutes per task against 75.
What the benchmarks say, and what they don’t
A benchmark is a fixed set of problems, run identically across different models so their scores can be lined up. OpenAI published three headline results for Astra: 98% on FrontierMath Tier 4, which tests advanced mathematical reasoning and which OpenAI’s own results table puts more precisely at 97.6%, 99.9% on ARC-AGI-3, which tests abstract problem-solving, and 100% on ExploitBench, which tests finding software vulnerabilities and is the reason OpenAI classifies this model at its highest cyber-risk tier. Reporting at launch placed it ahead of both GPT-5.6 Sol and Anthropic’s Claude Fable 5.
Those are OpenAI’s own evaluations on tests it selected, so read them as an indication of where the model is strong rather than a guarantee of what you’ll get, the same caution we applied to Anthropic’s figures when Claude Opus 5 arrived with its own benchmark claims. A score near 100% also tells you less than it appears to, because it usually means the test has run out of room rather than that the problem is solved.
Good to know: File creation, browsing and desktop control are three different capabilities. Producing a file is one: you ask for a comparison and get a spreadsheet to download, with nothing opening on your machine. Browsing is another: the model works through a website, picks options and fills in a form. It uses a browser of its own with a separate profile rather than the one you have open, so you choose what it can reach. Desktop control is the third and the one people picture: it opens an application on your computer and works the interface. They need different things switched on, so the one you want decides what you have to set up.
How to use GPT-6 Astra right now
There are three ways in, and almost everyone reading this wants the first.
- Through the ChatGPT app, which is what most people mean by using it. Astra is built into the standard ChatGPT website and the iPhone and Android apps, and you need a paid plan to reach it, since free accounts don’t have it. Plus, Pro, Business and Enterprise subscribers all do, and Pro, Business and Enterprise also get a heavier version called GPT-6 Astra Pro alongside it. Start a new chat, open the model dropdown, and you’ll find GPT-6 Astra listed under the Work or Codex modes alongside the GPT-5.6 models it replaces.
- To control applications on your own computer, there’s one more step. It runs through the ChatGPT desktop app rather than the browser version, has to be switched on before it does anything, and asks your operating system for permission to see and control the screen. On Windows it works on the active desktop, so you can’t carry on using that screen while it runs. Availability varies by region.
- Through the API, which is for software rather than people. An API is the connection that lets one piece of software talk to another, so this route is how a tool you own would send text to Astra and get an answer back without a person typing anything. This route applies only if custom software has been built for you, or by you, that sends requests to OpenAI directly, such as a chatbot serving your website visitors. The model name to use is gpt-6-astra, and it’s available from OpenAI directly as well as through Amazon’s AWS Bedrock and Microsoft Azure. A developer can also connect Astra to a browser or a desktop environment of their own, which is how computer use reaches software outside ChatGPT.
Keep in mind: Safety checks can pause a job part-way through. OpenAI says the extra monitoring around Astra can slow, pause or stop legitimate work. In ChatGPT or Codex a paused task asks you to review the action before it carries on, and through the API the task stops outright. On a long job that runs unattended, that means checking on it rather than assuming it finished.
How much does GPT-6 Astra cost?
The answer depends on how you reach it, and there are two quite different ways to pay. The 2.5x figure everyone is quoting applies to the per-token rates, not to your ChatGPT subscription (tokens are how AI usage gets counted, with one token working out at roughly three-quarters of a word).
If you use the ChatGPT app
You pay a flat monthly fee and Astra is included on the paid tiers, so choosing it instead of a cheaper model costs you nothing extra per message. What it costs you is headroom. Every plan has usage limits, and a heavier model works through that allowance faster, so the practical effect of switching to Astra for everything is that you hit your ceiling sooner in the day. Free accounts don’t reach it at all.
If you build on the developer side
The API route described above is where the 2.5x multiple bites, because you’re billed per token rather than by subscription. Two rates apply: a lower one on everything you send in, which providers call input, and a higher one on everything that comes back, called output. Roughly 750 words works out at 1,000 tokens.
| Model | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| GPT-5.6 Sol | $4 | $20 |
| GPT-6 Astra | $10 | $50 |
Sol’s $4 and $20 are promotional rates, running at least into late November 2026, so the 2.5x gap is measured against a discount with an end date. And requests above 272,000 input tokens move onto a second, higher rate card, which matters for the long-document work this model is built for.
Repeated material is where you can cut the bill. Anything the model has seen before can be cached, which costs $1 per million rather than $10, a 90% saving on background you send again and again.
What 2.5x means for a month of real work
A multiple is hard to act on, so here’s what it looks like against work a small business actually does. The figures below are illustrative calculations for text work, not measured computer-use sessions, converting text at roughly 750 words per 1,000 tokens.
If your billing is per token, these give you the rough shape of a bill rather than the bill itself. If you pay a flat monthly fee, they show you something else: which kinds of work draw hardest on the allowance you have already bought.
| A month of… | What that involves | On GPT-5.6 Sol | On GPT-6 Astra |
|---|---|---|---|
| Blog content | Four 1,200-word posts, each drafted from a brief plus background material, then revised twice | About $0.55 | About $1.40 |
| A product catalog rewrite | Descriptions for 200 products, around 150 words apiece, worked up from the specs you already hold | About $1.75 | About $4.35 |
| Social and email copy | A month of posting: twenty short social captions plus four campaign emails | Under $0.20 | About $0.40 |
| One long research job | Reading a 300-page report and producing a formatted summary | About $1 | About $2.40 |
Run all four in the same month and you spend around $3.40 on Sol against around $8.60 on Astra. The multiple is 2.5 and the difference is five dollars. For a business already paying a monthly subscription, the headline sounds alarming and the actual number is not.
Repetition on the long job is where it adds up. One 300-page report runs to $2.40, so working through twenty of them in a month comes to $48, several times the cost of everything else in that table combined. The volume you send in is what moves these numbers, more than the volume that comes back. One qualification, and it cuts against the headline. A rate is charged per token, and Astra uses fewer tokens than Sol on the same work, so the gap in what a finished job costs is smaller than 2.5x. On several of OpenAI’s own published comparisons Astra completed tasks at a lower estimated cost than Sol despite the higher rates, including around 43% lower on a CAD reconstruction test and around 27% lower on a scientific research one. So 2.5x describes the rate card rather than your bill, and what a job actually costs is something only your own work will tell you.
Pro tip: Before you agonize over model choice, look at what you’re sending. Brand guidelines, product lists and tone-of-voice notes pasted fresh into every conversation get charged again on each one, either against your bill or against your allowance. Load that material once into a saved project or an attached file the model can consult, rather than pasting it fresh each time. On the API side, repeated material may also qualify for cached pricing at a tenth of the standard input rate, though caching has its own eligibility rules and expiry, so treat it as a reduction rather than a free pass.
What is GPT-6 Astra best for?
Astra earns its price on long jobs that end in a finished thing. On short, routine work, try the cheaper model first and see whether the difference shows.
Long multi-step jobs that end in a deliverable
Put in terms of jobs sitting on your own list, that looks like:
- Research that ends in a document. Ask it to look into a topic, gather what’s published, and hand you a formatted brief rather than a wall of notes you then have to organize yourself.
- Building a spreadsheet from scattered sources. Three suppliers’ public pricing pulled into one comparison table, with the columns you asked for and the arithmetic already done.
- Work that runs through stages. Going through your published pages, working out which ones still earn their place, and turning that into a plan, with the model holding the brief across all three steps rather than losing it between them.
- Following your templates. OpenAI says it matches your existing templates and your writing and visual style, and pulls in only the context a job needs rather than repeating everything it was given. For recurring reports that’s the difference between a draft and something you can send.
We ran our own test and gave GPT-6 Astra and the model it replaces, GPT-5.6 Sol, the same prompt:
“Research the current published pricing for Mailchimp, Brevo and MailerLite for a business with 5,000 email subscribers sending around 20,000 emails a month. Build me a comparison table showing the monthly cost on each plan that fits those numbers, what’s included at that tier, and any limits I’d hit. Then tell me which one you’d pick for a small business at that size and why. Give me the table as something I can download.“
Both returned a working spreadsheet and both landed on the same recommendation for the same reason. Where they parted was how much of the picture came back.
Sol gave one plan per provider and explained the logic underneath.
Meanwhile, Astra laid out every tier from all three, stated the assumptions it had priced against, and flagged two catches in the fine print that Sol passed over. Both were usable. Only one showed the full ladder of plans, which matters if your list grows.
When the cheaper model is the right call
A frontier model is overkill for a good share of the week. Captions, a reply to a customer question, notes from a meeting turned into something readable, sorting and labeling, the first pass at something you write regularly: hand any of that to the cheaper model and you’ll be hard pressed to find fault with what comes back. That’s a large share of the marketing work AI has taken over, and none of it needs a frontier model. Save Astra for the work where the difference is visible.
Is GPT-6 Astra AGI?
OpenAI president Greg Brockman called Astra a generational leap, said he personally believes OpenAI has reached artificial general intelligence, and closed his launch briefing with “Welcome to the AGI era,” while leaving users to decide whether Astra meets the definition. Artificial general intelligence, usually shortened to AGI, is the idea of a system that can handle any intellectual task a person can, rather than being strong in particular areas.
Independent researchers are more cautious. Toby Walsh, professor of AI at the University of New South Wales, said OpenAI is clearly neck and neck in the race, and that the intelligence in artificial intelligence is still today very jagged, meaning it can be superhuman at one task and fail at something a child manages.
Something interesting: Astra is the first OpenAI model where other AI models helped supervise its training. OpenAI built it on its largest training run to date, using more than 100,000 GPUs at its Stargate site in Texas, and says other models played a significant part in overseeing that process. It doesn’t change what the model does for you, and it’s a shift in how frontier models get made: the systems are now involved in building their successors.
For your business, this changes nothing about what you should do this month. Whether Astra clears a definition researchers have argued over for decades makes no difference to whether it writes your product descriptions better than the model you’re paying for now. That question has an answer, and you can get it in an afternoon by running both.
The bottom line: Is GPT-6 Astra worth it for your business?
Nothing about this launch overturns the arrangement most small businesses already have: something cheap running the daily work, and something stronger held back for jobs that are long, layered, or end in a finished artifact. The difference this time is that the stronger option carries a premium, which turns the question from whether Astra performs better into whether the work you’d give it justifies paying 2.5 times as much.
No write-up settles this, ours included. Pick something from your actual workload, put it through whichever model you’re paying for today and then through Astra, and set the two finished outputs against each other along with what each one consumed. A couple of jobs is usually enough to sort your work into two piles, and that sorting is worth more than any published score.
Running that kind of comparison is easier when the models sit together rather than behind separate logins and separate invoices. SiteGround AI Studio carries leading models from OpenAI, Anthropic and Google on one subscription, chosen from a dropdown at the start of each chat. That makes running the same question on two models a matter of changing one setting.
[97]
FAQs
GPT-6 Astra is OpenAI’s frontier AI model, released on September 3, 2026. OpenAI describes it as its most intelligent and aligned model, built less like a chatbot and more like a computer operator that can use software, browse the web, produce documents and carry out long multi-step tasks.
GPT-6 Astra is priced at 2.5 times its predecessor, GPT-5.6 Sol, on the per-token developer rates. Through the ChatGPT app it’s included in the paid subscription tiers rather than charged separately, so the cost shows up as usage limits rather than as a bill.
Not settled, and the people best placed to know disagree. OpenAI president Greg Brockman said he personally believes OpenAI has reached artificial general intelligence and that it might be this model, while leaving users to decide. Independent researchers are more cautious, describing current AI capability as still very jagged.
Mostly it does the same kinds of things better and faster. GPT-5.6 Sol already handled computer use, and OpenAI scores both models on the same tests. Astra is reported as state-of-the-art on computer use, browsing, software engineering and professional work, finishing computer-use tasks in roughly 47% less time, and scoring 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.
It depends on the work. A model priced at 2.5 times its predecessor earns that on long multi-step tasks that end in a real deliverable, such as a researched document or a built spreadsheet. For short marketing copy and quick questions, a cheaper model produces the same result.
Yes, in some form. GPT-6 Astra rolled out to paid ChatGPT tiers and through the OpenAI API, so free accounts don’t reach it. You can also reach current leading models through a bundled workspace such as SiteGround AI Studio, which includes several models on one subscription.
It means the model can work the way a person works at a screen rather than only writing text. OpenAI describes it using applications, browsing the web, inspecting what’s displayed, building websites, producing documents and spreadsheets, and running quality checks on its own output.
Not for attacking anything, though security work isn’t off the table. OpenAI classifies GPT-6 Astra as reaching its Critical cybersecurity threshold, and says defenders can use the public version for work such as secure code review and patching. What it refuses is the advanced end, including writing proof-of-concept exploits. Less restricted access is limited to vetted organizations through an application-based program.
