GPT-6 Astra: What OpenAI Announced and What It Means for Your Business

On September 3, 2026, OpenAI introduced GPT-6 Astra and called it “the world’s most intelligent and aligned model.” OpenAI president Greg Brockman went further: “Welcome to the AGI era.” Behind the headline are two things that actually matter if you run a company: it is the first model OpenAI has classified at the Critical cybersecurity threshold under its own preparedness framework, and it costs $10 and $50 per million input and output tokens. It is neither magic nor vapor. It is a real jump, an expensive one, with fine print almost nobody is reading.

Here is the announcement video:

The 60-Second Version

• What it is: GPT-6 Astra, OpenAI’s new flagship, built around computer use: navigating interfaces, filling spreadsheets, operating software the way a person would.
• When: announced September 3, 2026. It shipped first to a limited set of organizations, and “over the coming days” reaches all ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API and AWS.
• What it costs: $10 per million input tokens and $50 per million output tokens in standard mode. Fast mode runs roughly double the price at up to 2.5x the speed.
• Memory: a 1 million token context window, roughly 1,500 pages.
• The red flag: it is the first model to cross the “Critical” cybersecurity threshold in OpenAI’s Preparedness Framework. It can find and exploit previously unknown vulnerabilities without step-by-step human guidance.
• The fine print: that 99.9% on ARC-AGI-3 relies on a specialized harness. Called through the API the normal way, the same model scores between 17% and 63%.

What Makes Astra Different

Earlier models were, fundamentally, very good writers. You asked for a document, an analysis, or a block of code and got it. Wiring it into your systems meant somebody had to build an integration for every tool.

Astra is trained to use the computer: open a browser, search Google Maps, fill out a contact form, move around inside a spreadsheet, chain twenty steps without losing the thread. The practical difference is large. You used to need an API for every system you wanted to automate. Now, in theory, the model can drive the system through the same screen your staff drives it through.

The analogy: earlier models were a brilliant consultant on the phone. You explained the problem, they dictated the answer, and somebody else had to type it in. Astra is that same consultant sitting at the keyboard. That is a category jump — and the reason the safety conversation got serious.

The Launch Numbers

What the Benchmarks Actually Mean

This is where the enthusiasm needs a governor. Three honest readings:

1. The headline number carries an asterisk. That 99.9% on ARC-AGI-3 depends on a stateful harness built by OpenAI that lets the model accumulate context across attempts. If your team calls the API the way teams normally call APIs, the same model lands somewhere between 17% and 63% on that test. It is not a trick, but it does mean the score you buy is not the score in the press release. What your implementation delivers depends as much on the engineering around the model as on the model.

2. It is not the best at everything. On Humanity’s Last Exam with tools, Astra comes in below Claude Fable 5.1 (57.2% vs 65.0%, per DataCamp’s analysis). “World’s most intelligent” is marketing; the reality is there are still tasks where another model wins. We worked through that comparison in Claude Opus 5 vs Fable 5 and GPT-5.6 Sol vs Claude Fable 5.

3. The benchmark that measures real work is missing. GDPval, OpenAI’s own test of performance on real professional tasks across 44 occupations, was absent from the launch materials. It is the metric closest to “does this help in my office,” and it is the one that did not ship. Worth remembering before signing a contract based on a press release.

The number that is solid and does matter: OSWorld 2.0, which measures actual computer operation. 72.6% against 65.7% for the previous model, finishing tasks in roughly half the time. That is the kind of gain you feel in operations, not on a chart.

Grounded Costs

At $10 and $50 per million tokens, here is the math for an agent that reads documents and writes a summary or a data entry. These are our estimates from the published rates, not OpenAI figures:

The practical takeaway: for high-value, discrete tasks — reviewing contracts, auditing inventory, closing the books — the cost is irrelevant next to the hours saved. For high-volume, low-value repetition, Astra is an expensive model and you are probably better served by a cheaper one. The right question is not “what does a token cost” but “what does the task cost, and what does doing it by hand cost me today?”

What It Means for a Small Business

1. Automation Stops Requiring an Integration for Everything

If your supplier’s system has no API, that used to mean it could not be automated without custom development. A model that drives the screen changes that equation for medium-volume processes. Careful, though: it changes the equation, it does not erase it. An agent filling forms on somebody else’s portal is fragile — the portal changes and it breaks. A real integration is still better when the process is critical, which is where proper business process automation on your own systems earns its keep.

2. The Cybersecurity Part Is Not a Footnote

OpenAI classified Astra at the “Critical” cyber capability level: it can find unknown vulnerabilities and build exploit chains on its own. The company restricted those capabilities at launch and added safeguards following the Hugging Face incident. Translation for your business: the same capability that makes an agent useful makes it dangerous when somebody else points it at you. If you have been putting off the basics — two-factor authentication, tested backups, least-privilege access — this is the reminder. We covered the concrete case in AI Agent Cyberattacks: The Hugging Face Lesson.

3. Customers Will Find You by Asking an AI

Every generation of these models pushes more searching into the chat window and less into the list of blue links. If your business does not surface when someone asks ChatGPT for a provider in your trade and your city, that sale never existed for you. It is a different job than classic SEO, and it deserves its own plan.

4. What Does Not Change

Astra does not know your inventory, cannot see your purchase orders, and has no idea which customer owes you money. A smarter model does not fix messy data — it processes it faster and gets things wrong with more confidence. If your information lives in three spreadsheets and one manager’s head, the model is not your bottleneck. Where agents do earn their keep is covered in AI Agents for Business.

How to Decide: A Two-Week Pilot

1. Pick one task that is repetitive, measurable, and low-risk. Invoice entry and contract review are the two classics that work.
2. Measure what it costs today: hours per week times that person’s real cost. Without that number there is no way to know whether you won.
3. Try the model you already pay for before jumping to Astra. On plenty of tasks a cheaper model delivers the same result, and that isolates whether the problem was the model or the process.
4. Run 50 real cases and count the errors by hand. Ninety percent accuracy sounds great until you price the other ten percent.
5. Decide on cost per task, not on token price and not on the benchmark in the press release.
6. If the agent will touch outside systems, define what it may do alone and what needs human approval. Write it down before you switch it on.

FAQ

Can I use GPT-6 Astra yet?

It depends. On September 3 it went to a limited set of organizations, and OpenAI said it would reach all ChatGPT Plus, Pro, Business, and Enterprise users over the following days, along with the API and AWS. If you are on a paid plan, it is a matter of days.

What does it cost?

Through the API, $10 per million input tokens and $50 per million output in standard mode; Fast mode is roughly double. Inside ChatGPT it draws on your plan’s allowance, with additional credits available for purchase.

Is it really AGI?

Brockman said “welcome to the AGI era” while acknowledging the definition is ambiguous. A model that scores 100% on an exploitation benchmark and trails a competitor on a reasoning benchmark is not a general intelligence; it is a system that is extraordinarily good at a set of tasks. For your business, the useful question is not whether it is AGI but which specific task it solves and at what price.

Should I switch AI vendors over this?

Not on the strength of a press release. Leadership among these models changes every few weeks, and the real cost of switching sits in the work around the model, not the model itself. Build processes you can move between vendors rather than marrying whichever name is on top this month.

Does this make my business less secure?

Offensive capability rises for everyone, attackers included. OpenAI restricted the most sensitive functions at launch, but the direction is clear. The sensible response is not panic — it is having the digital hygiene basics current, which is exactly what most small businesses have been deferring.

Sources

• Axios: OpenAI releases GPT-6 Astra, says it may represent AGI
• VentureBeat: “Welcome to the AGI era” — benchmark and pricing analysis
• CNBC: OpenAI announces rollout of GPT-6 Astra
• DataCamp: GPT-6 Astra features, benchmarks and pricing
• Artificial Analysis: GPT-6 Astra price and context window
• 9to5Google: launch and phased availability

If this launch has you wondering which of your processes could survive an AI agent and which would break, that is the conversation we have every day. Book a 30-minute call and we will tell you straight which processes to automate first and which ones to leave alone.