OpenAI dropped GPT-6 Astra today. For the first time in a long while they are ahead of Claude on the benchmarks, so I opened the announcement page cold and read the whole thing on camera.
The scores are impressive. They are also not the part that should change how you run your business this month.
The numbers, off their own page
Everything here is theirs. I did not test it, and I will get to that.
They are calling Astra state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work. It saturates FrontierMath Tier 4 at ninety eight percent, having already solved open problems in mathematics. It hits ninety nine point nine on ARC AGI 3 and a hundred percent on ExploitBench.
Against Claude Fable 5.1, which came out three days ago and which I have been using every day since, Astra is clearly out in front on their charts. That has not happened in a while. Credit where it is due.
The other line that matters is price. On several of these it is close to Claude on accuracy and roughly half the cost per output token. Close but much cheaper is a real product decision, not a rounding error.
Watch their demo, not their charts
Here is the part I actually rewound.
In their own demo, somebody stands there and gives it six jobs in one breath. Build me a 3D game where I dodge asteroids with the arrow keys and boost with space. Also I am hungry, order the beef and rice from that place I used last week. My law firm needs a licensing agreement template, draft one for the lawyers to look at. Find me a tennis court in the Lower Haight this afternoon. Book the five o'clock. Now make me an STL file I can send to my 3D printer.
Then they walk away and it does all of it.
The technical achievement is real. But the thing to notice is how they talked to it. They were not asking it for help. They were telling it what to do and leaving.
That is the entire gap between the people getting value out of AI right now and the people who tried it, got something mushy back, and decided it was overhyped. Same tool. Different sentence.
Computer use is the line item for a small business
Strip out the math benchmarks and look at what they say it does. Filling out online forms. Updating customer records in a CRM. Organizing a calendar. Formatting a legal document. Filling a 1040 from your W-2s. Apartment hunting. DMV appointments.
Read that list again as a business owner instead of as a developer. That is your Tuesday. That is the pile of small, tedious, must-be-done-correctly work that eats a person's afternoon and that you cannot really hire out cheaply because it touches your actual customer data.
Their accuracy numbers on OSWorld, which is the benchmark for driving a real computer, went up meaningfully and the cost went down by close to half. A model that can reliably drive a browser and a CRM is worth more to a service business than a model that can solve open problems in mathematics.
I already felt this before the announcement, honestly. Their desktop app has been getting good. I built a client's Shopify store with it last week, changes documented and screenshotted, and it handled work my team had bounced back to me.
What I have not done, so you know where I stand
I have not used Astra. I am out of tokens on my subscription and it is not fully rolled out yet. This is me reading the research the day it landed and telling you what I think it means, not a review.
I am saying that plainly because the internet is about to fill up with people reviewing a model they have not touched. Every number above is OpenAI's, published on their own page, and benchmarks are marketing until somebody independent reproduces them.
What a San Jose business should actually do about this
Probably nothing this week. That is the honest answer.
Do not rip out what you are running because a chart moved. I use Claude for most of my build work and I am not switching on a launch day. The models will trade places again in a month, and then again after that. If your business plan depends on which model is on top, your business plan is a coin flip.
What is worth your attention is the habit underneath the demo. Practice telling instead of asking. Give it the whole job, with your real constraints and your real details, and then let it work. Most local business owners I sit down with in San Jose are still typing one line questions into a chat box and judging the entire technology by what comes back.
The second thing is to figure out which parts of your week are the boring computer work. The form filling. The record updating. The quote copied from one system into another. Write those down. That list is your actual automation roadmap, and it does not change when a new model ships. The models just keep getting better at the same list.
Every one of those tasks touches customer data, so the same rule applies as everywhere else in my shop. A person looks at the output before it counts. Approval steps are not a lack of ambition, they are what makes the thing safe to leave running.
Watch the whole thing
I read the announcement page start to finish on camera, including their demo and the cost charts.
Watch: OpenAI Just Dropped GPT-6 Astra and It Beat Claude Fable 5.1
If you want to learn how to talk to these things properly, that is free in the Academy, live on Tuesdays and Thursdays, prompts included.
And if you want the boring computer work in your business actually handled, that is the job. On Point Tech Solutions builds systems and automation for businesses across San Jose and the Bay Area. Tell us what is eating your week and we will tell you honestly whether AI is the right tool for it yet.