GPT-6 Astra: What OpenAI's Agent Model Means for Your Business

OpenAI released GPT-6 Astra on 4 September 2026. The benchmarks are strong. That is not the part that should interest you.

The part that should interest you is this: GPT-6 Astra can open a browser, read a page, compare options, fill in a form, and book something on a customer's behalf. Your next customer might never look at your website. Their assistant will.

TL;DR

• GPT-6 Astra scored 72.6% on OSWorld 2.0, a test of whether an AI can operate a computer, against 65.7% for GPT-5.6 Sol and 70.2% for Claude Opus 5.
• It finished those tasks in about 40 minutes each, against about 75 minutes for GPT-5.6 Sol, which is roughly 47% less time per task.
• Astra scored 99.9% on ARC-AGI-3, an abstract reasoning test, where GPT-5.6 Sol scored 7.8% and Claude Opus 5 scored 30.2%.
• API access costs $10 per million input tokens and $50 per million output tokens, under the model name gpt-6-astra.
• Astra does not lead everywhere. Claude Fable 5.1 scored 65.0% on Humanity's Last Exam with tools, against 57.2% for Astra.

All figures here are OpenAI-reported, from OpenAI's own announcement.

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's flagship AI model, released on 4 September 2026. OpenAI calls it its most intelligent and most aligned model to date. It is built to use a computer the way a person does: browsing websites, filling in forms, and completing multi-step tasks, rather than only answering questions in a chat window.

It is rolling out to a limited set of organizations first, then to ChatGPT Plus, Pro, Business and Enterprise over the following days. Developers can reach it through the API as gpt-6-astra, and it is also available on Microsoft Azure and AWS Bedrock.

OpenAI reports state of the art results on computer use, browsing, software engineering, cybersecurity, science and professional work. It also says Astra helped establish a new result on gaps between prime numbers, improving a bound on short prime gaps from 240 to 186.

The benchmarks, in plain English

Benchmark names are jargon. Here is what each one actually measures, and how Astra did.

Coding

Terminal-Bench 4.0 measures how often a model finishes a real command-line software task. Astra leads, but the gap over Anthropic's models is narrow.

The wide gap is against GPT-5.6 Sol (37.3%) and Gemini 3.8 Flash (19.1%). The narrow gap is against Claude Fable 5.1 at 55.8%. On DeepSWE v1.1, another coding test, the field is tighter still: Astra 74.1%, Gemini 3.8 Flash 73.8%, Claude Opus 5 73.7%, GPT-5.6 Sol 72.7%.

Computer use

This is the category that matters for your business. Both of these tests check whether a model can drive a real computer to finish a job.

Astra also scored 92.7% on ScreenSpot-Pro without tools, a test of whether a model can correctly locate things on a screen. GPT-5.6 Sol scored 76.9%. Claude Fable 5 scored 87.3%, though OpenAI notes that figure came from Mythos, the less-safeguarded version of that model.

Speed

Accuracy is only half of it. A booking agent that takes 75 minutes to fill in a form is not a booking agent anyone will use.

OpenAI also reports that Astra completes tasks 1.9 times faster than GPT-5.6 Sol on Mind2Web when it runs inside the updated Codex agent scaffolding.

Reasoning and memory

ARC-AGI-3 tests abstract reasoning on puzzles the model has not seen before. Astra scored 99.9%. GPT-5.6 Sol scored 7.8% and Claude Opus 5 scored 30.2%. That is not an incremental step.

On MRCR v2, which tests whether a model can find specific facts buried in a very long document, Astra scored 96.3% across 512K to 1M tokens. GPT-5.6 Sol scored 73.8%. On GPQA Diamond, a set of graduate-level science questions, Astra scored 96.0% against 95.3% for Gemini 3.8 Flash.

The real headline: AI that uses a computer like a person

Read the demos OpenAI chose to show, because they tell you more than the scores do.

Finding a pediatrician. Hunting for an apartment. Booking a DMV appointment. Filling in forms. Updating CRM records. Organizing calendars.

Those are not developer tasks. They are errands. They are the exact things your customers currently do on your website, by hand, one click at a time.

Astra scored 59.3% on Agents' Last Exam and 41.4% on AutomationBench, which measures multi-step office work. It scored 91.5% on BrowseComp, a test of finding hard-to-locate information on the open web. It is not perfect. It is good enough to be useful, and it is getting faster.

So the question changes. It used to be "can a customer find my website?" Now it is also "can a software agent read my site, understand my hours and prices, and complete my booking form without a human?"

If the answer is no, the agent will pick a competitor whose site it can read. It will not send you an email explaining why.

Where Astra does not win

Any article claiming a clean sweep is selling something. Astra loses in several places, and OpenAI publishes those results too.

Claude Fable 5 also scored 63.8% and Claude Opus 5 scored 63.6% on Humanity's Last Exam, both ahead of Astra. On the two Artificial Analysis indices, Claude Fable 5 sits at 62.1 and 67.2 respectively.

There is a pattern here. Astra leads decisively on doing things: operating a computer, browsing, finishing tasks. It does not lead on every measure of raw knowledge. For a local business, the first category is the one that changes your week.

One more caveat worth reading. OpenAI notes that some Claude comparison scores came from modified evaluations, so treat cross-vendor numbers as indicative rather than exact.

Safety and cybersecurity

OpenAI put real work into the safety side, and the numbers support that.

On an internal test where a model was given an impossible task, GPT-5.6 Sol went beyond its authorized scope 48% of the time. Astra did so 0% of the time. On OpenAI's internal hallucination benchmark, Astra scored 4.2% against 12.2% for GPT-5.6 Sol.

OpenAI says Astra meets its "Critical" cybersecurity threshold under its Preparedness Framework. It ships with extra safeguards, and advanced offensive security tasks are refused at launch.

OpenAI is also candid about one weakness. In testing, Astra's written reasoning was harder to monitor than GPT-5.6 Sol's. A model whose thinking is harder to audit is harder to catch when it goes wrong, and OpenAI says so plainly rather than burying it.

Pricing and availability

Zero Data Retention is supported for eligible API customers, which matters if you work in health, legal or finance.

What this means for your business

Picture a real request. Someone opens ChatGPT and types: "find a pediatric optometrist near me with Saturday hours and book me in."

The agent searches. It opens five or six websites. It reads them, compares hours and services, picks one, and fills in the booking form.

Every step of that is now a ranking factor. Not "ranking" the way Google means it. Ranking the way a shortlist works.

Think about what the agent needs at each step. It needs to find you, which is the old problem. Then it needs to confirm you are a pediatric optometrist and not a general one. Then it needs your Saturday hours as a fact, not a sentence. Then it needs a form it can actually fill in.

Fail any one of those and you drop out. There is no partial credit, and there is no second page of results for the customer to scroll.

That is the real shift. Search gave you a chance to be chosen by a person who could squint at a badly built page and work it out. An agent will not squint. It moves on.

Here is what I would do about it, in order.

1. Put your facts in crawlable HTML

Hours, services, prices, service area, booking link. In the page source, as text.

If your hours live in an image, a PDF, or a widget that loads after JavaScript runs, an agent may not see them. Many AI crawlers do not run JavaScript. A page that looks complete in your browser can be nearly empty to a bot.

There is a quick way to check this yourself. Open your site, right click, and choose "view page source." That raw text is roughly what a crawler gets. If you cannot find your phone number or your hours in it, neither can they.

Watch for vague copy too. "Evenings and weekends by appointment" reads fine to a person and means nothing to software comparing you against a clinic that published "Saturday, 9am to 2pm." Be specific enough to be matched against a query.

This is the least glamorous fix on the list and the one that moves the needle most. We wrote about the same problem from the transaction side in our guide to agentic commerce and machine-readable data.

2. Add schema markup that names your entity

Schema markup is structured data that tells a machine what your business is, not just what your page says.

An agent comparing four optometrists does not want to parse your homepage prose. It wants LocalBusiness, openingHoursSpecification, areaServed, and a clean sameAs list. Give it those and you stop being a guess.

Our JSON-LD schema templates for AI cover the exact markup, and The AI Visibility Engine explains how we sequence this work.

3. Make your booking form completable by a machine

Open your own booking page. Count the steps. Now imagine software doing it.

A three-field form with clear labels is completable. A form that requires a phone call, a login, a CAPTCHA on every attempt, or a date picker that only responds to a mouse drag is not. If your funnel ends in "call us to book," an agent cannot finish the job.

This is where local businesses lose quietly. A dental clinic with online booking gets picked over one without. So does a law firm with a plain intake form, an optometrist whose Saturday hours are in the markup, and a home services company that publishes a service area instead of "the GTA."

Reviews still matter, but not the way most owners think. We covered why in Google reviews mean nothing to ChatGPT. You can see the citation results this work produces on our client results page.

What not to panic about

You do not need to rebuild your website. You do not need an app, a chatbot, or a contract with OpenAI.

Most of the work above is copy and markup on pages you already own. A clinic with clear hours in its HTML and working schema is in better shape than a competitor who spent thirty thousand dollars on a redesign that renders entirely in JavaScript.

You also have some time. Astra scored 41.4% on AutomationBench, so agents still fail plenty of multi-step office tasks. That number will climb. The businesses that prepare while it is climbing are the ones that get picked when it stops being a novelty.

And none of this replaces the basics. An agent that shortlists you still hands a human the final decision, and that human reads your reviews. Being machine-readable gets you onto the list. Being good keeps you there.

The businesses that win the next two years will not be the ones with the prettiest websites. They will be the ones a machine can read, compare and complete a booking on.

Want to know how you look to an AI agent right now? Run a free AI visibility scan. It checks your site the way ChatGPT, Claude and Gemini do, and tells you what they can and cannot see.

Frequently asked questions

What is GPT-6 Astra? GPT-6 Astra is OpenAI's flagship AI model, released on 4 September 2026. OpenAI describes it as its most intelligent and most aligned model so far. Unlike a standard chatbot, it is built to operate a computer: browsing the web, comparing options, filling in forms and completing multi-step tasks on a person's behalf.

Is GPT-6 Astra better than Claude? It depends on the task. Astra leads on computer use, scoring 72.6% on OSWorld 2.0 against 70.2% for Claude Opus 5. Claude leads on some knowledge tests: Claude Fable 5.1 scored 65.0% on Humanity's Last Exam with tools, against 57.2% for Astra. For agent work, Astra is ahead. For hard expert questions, Claude often is.

How much does GPT-6 Astra cost? Through the API it costs $10 per million input tokens and $50 per million output tokens, under the model name gpt-6-astra. A fast mode offers up to double the speed at double the price. For everyday use it is included in ChatGPT Plus, Pro, Business and Enterprise plans at no extra charge.

Who can use GPT-6 Astra? OpenAI rolled it out to a limited set of organizations first, then to ChatGPT Plus, Pro, Business and Enterprise subscribers over the following days. Pro, Business and Enterprise plans also include GPT-6 Astra Pro. Enterprise administrators must switch it on, because it is off by default. Developers can also reach it through Microsoft Azure and AWS Bedrock.

Can GPT-6 Astra book appointments or fill out forms? Yes. OpenAI's own demonstrations include searching for a pediatrician, apartment hunting, booking a DMV appointment, filling in forms, updating CRM records and organizing calendars. It scored 72.6% on OSWorld 2.0, which tests real computer operation, and finished those tasks in roughly 47% less time than GPT-5.6 Sol.

Does GPT-6 Astra change SEO? It adds a second audience. Your site now needs to be readable by software agents as well as people. That means facts in crawlable HTML rather than images or JavaScript widgets, schema markup describing your business, and booking forms simple enough for an agent to complete without a phone call.

Sources

• OpenAI: GPT-6 Astra announcement

All benchmark figures in this article are OpenAI-reported, taken from that announcement and its accompanying footnotes. Cross-vendor comparisons should be read as indicative: OpenAI notes that some competitor scores came from modified evaluations, and that one Claude ScreenSpot-Pro figure came from a less-safeguarded model variant.