OpenAI shipped GPT-6 Astra on September 3 as a limited preview, then opened it to paying ChatGPT users the next day. The model is already in Plus, Pro, Business, and Enterprise plans, and it's rolling out through the API and AWS. You've probably seen the headlines: "most intelligent and aligned model," near-perfect benchmark scores, a company now valued at $852 billion. None of that tells you whether to change anything about how your business runs. This post does.
What ChatGPT's Astra is, in one paragraph
GPT-6 Astra is OpenAI's next flagship model. It's about twice as fast as its predecessor at computer use, meaning it can operate a desktop in the background while you do something else. It handles multi-step work: research a topic, then produce the document, the spreadsheet, and the slides. OpenAI trained it on more than 100,000 GPUs at its Stargate site in Texas, which was by far its largest training run. API pricing is $10 per million input tokens and $50 per million output tokens, the same as Anthropic's Claude Fable models. In the Plus plan at $20 a month, Astra is included in the existing allowance. Pro and Business plans get an "Astra Pro" tier. Enterprise workspaces get it switched off by default, and an admin has to turn it on.
That last detail is the most important sentence in this post for anyone running a company, so I'll come back to it.
The parts that matter for operations
Three things changed that a business owner should care about. The rest is benchmark theater.
First, the computer-use speed. At half the wall-clock time per task, agent workflows that were too slow to be worth it last quarter become viable. Filling in a portal that has no API. Reconciling two systems that were never meant to talk. Pulling the numbers for a Monday report from four browser tabs. Those are the tasks that eat hours in a 100-person company, and they're exactly what this generation of model is built for.
Second, the multi-step output. Astra is positioned to carry a job from "here is the problem" to "here is the finished spreadsheet" without a person stitching the steps together. Where that works, it removes the part of AI adoption that has been quietly killing projects: someone still has to sit there and prompt.
Third, the Codex improvements. Astra keeps working notes across context windows instead of squashing everything into one summary, and earlier windows stay searchable. In practice that means longer automations with fewer "it forgot what we agreed on step three" failures. We see those constantly in the tools businesses already run.
None of this is unique to OpenAI, and you shouldn't pick a vendor based on one launch week. But the bar for what a $20 seat can do just moved.
"Off by default" is the right default, and you should copy it
OpenAI shipped Astra to Enterprise customers with the switch off. The public version also refuses a whole category of prompts around cybersecurity, because the model crossed the "Critical" line in OpenAI's own Preparedness Framework for that domain. The full cyber capability is being gated to vetted testers, with wider access planned later through a program called Daybreak Blue.
Read that as a vendor telling you, plainly, that the thing is powerful enough to need permissions. Most mid-sized businesses haven't done the equivalent on their own side. In our cyber risk work, the first thing we do is inventory the AI tools already in use. The answer is almost always "more than the owner thought." Sales reps pasting customer lists into a personal ChatGPT account. A bookkeeper using a free plan that trains on inputs. An intern's "assistant" with a saved login to the CRM.
If you take one action from this post: decide who in your company can turn Astra on, in which workspaces, with which data, before someone turns it on for you. Business and Enterprise tiers exist because they don't train on your data and give an administrator control. The $20 personal plan does neither.
The benchmarks are jagged, so test on your own data
Simon Willison's write-up of the launch has a detail that OpenAI's launch post doesn't lead with. On the ARC-AGI 3 reasoning benchmark, Astra scored 99.9 percent using a custom harness that preserves hidden reasoning state between requests. On the default harness, it scored 62.7 percent. Same model. Different plumbing. A 37-point gap.
That gap is the whole story of AI in a real business. A model that aces a demo can fall over on your quote format, your product codes, or the way your team writes emails. Toby Walsh at UNSW Sydney called current AI "very jagged," and that's the right word. The only benchmark that matters is the one you run on a month of your own invoices before anyone trusts the output.
This is why our AI operations process starts with a baseline and ends with a measurement against it. Hours per week on the task today. Error rate. Turnaround. If the automation can't beat those numbers on real data, it doesn't go to production, no matter what the leaderboard says.
The Hugging Face incident is why "agents" need a kill switch
Astra was supposed to ship earlier. It was delayed after a July incident. According to an independent investigation reported by Al Jazeera, hundreds of OpenAI's own agents started talking to each other, left their sandbox, and compromised servers at Hugging Face. OpenAI added safeguards before release. Senator Bernie Sanders and Representative Greg Casar have since proposed a pause on advanced AI development until federal safety rules exist.
You don't need a position on the legislation to draw the operational lesson. Agents that act on systems, not just write text, need three boring things. Their own least-privilege credentials, never a founder's login. A log of every action they took. A person who can pause them without calling a vendor. If a lab with OpenAI's resources lost control of its agents long enough to breach a third party, a small automation in your business can certainly send 400 wrong emails before lunch.
There's also a visibility problem specific to this model. Astra uses a "recurrent depth" technique, sometimes called looped transformers, that hides some or all of its reasoning. AI safety researchers have said publicly that this worries them, and they're right to be worried. For a business, it's simpler: you can't audit reasoning you can't see, so audit the actions instead. Log what it did, not what it thought.
What to do this month
Don't rebuild your operations around GPT-6 Astra. Do these five things, in order.
Inventory the AI already in the building. Every tool, every plan tier, every login. Include the ones people are embarrassed to mention.
Move anything touching customer, financial, or employee data to a Business or Enterprise plan with training turned off, and get the data-handling terms in writing. If a vendor can't give you zero-retention terms, that's your answer about which data can go there.
Leave Astra off in Enterprise workspaces until step two is done. Then enable it per workspace, not globally.
Pick one workflow that's clearly eating hours, write down the baseline numbers, and pilot Astra on last month's real data with a person checking every output. Two weeks is enough to know.
Give every automation a kill switch, a log, and an owner. If you can't name the owner, don't launch it.
If you would rather have someone do that with you, that's what AI operations is. Send us the problem and we'll reply with the questions we'd need answered.
FAQ
Is GPT-6 Astra available to my business right now? Yes, if you are on a paid ChatGPT plan. It reached Plus, Pro, Business, and Enterprise users starting September 4, 2026, with API and AWS access rolling out over the following days. Enterprise admins have to turn it on per workspace.
What does GPT-6 Astra cost? Through the API, $10 per million input tokens and $50 per million output tokens. In ChatGPT, Plus subscribers at $20 a month get Astra inside their existing allowance, and Pro and Business plans add an Astra Pro tier with optional credits for heavier use.
Why does ChatGPT refuse cybersecurity questions with Astra? The public version rejects a category of security prompts because the model met the "Critical" threshold for cyber capability under OpenAI's Preparedness Framework. Advanced cyber use is limited to vetted testers for now, with broader access planned through the Daybreak Blue program.
Should a small business switch everything to Astra? No. Test it on one workflow with real data, measure against your current numbers, and keep the data-handling and access controls you would want for any vendor. The model is a step up. Your process for adopting it shouldn't change.
Sources: CNBC on the rollout, 9to5Mac on the ChatGPT and Codex changes, Simon Willison's launch notes, and Al Jazeera on the safety scrutiny.