OpenAI President Says It Reached AGI with Astra - DTNS 5347

**Tom Merritt** (0:00)
Don't hit it, no violence is necessary.
This is the Daily Tech News for Friday, September 4th, 2026, the 4th of September. We tell you what you need to know, give you important context and help each other understand.

**Jenn Cutter** (0:22)
Today, OpenAI says we have reached AGI. Oh, and it also appears another rogue agent hacked another site in Germany.

**Tom Merritt** (0:32)
Yeah, I mean, you know, it's just the world we live in. I'm Tom Merritt.

**Jenn Cutter** (0:35)
I'm Jenn Cutter.

**Tom Merritt** (0:37)
Let's start with what you need to know with those big stories.
Yeah, so we knew Astra was coming. OpenAI has been talking about how it reached this critical security thing and all of that. And now some people have it, not me. OpenAI released a new model called GPT-6 Astra to partners in its Daybreak program. That is the program that gives its most capable models to trusted partners first as a bit of a safety measure, while it continues to fine-tune the guard rails against misuse. It has guard rails. Even the Daybreak people have to have the guard rails, but they just want to make sure before they give it to the unwashed masses. Although I took a shower today, so I'm just masses. Anyway, a version with security guard rails for us is planned as long as you want to fork over and pay. Paid subscriptions will get it in the quote coming days. Both versions, the Daybreak version and the one that the rest of us will get, have security measures developed after they determine that it qualified as critical in its ability to discover zero-day exploits. In fact, Astra received a perfect score on exploit bench, which if you're not thinking too hard about it, sounds like a good thing. Oh, perfect score. Good job, Astra. That means that it can be used to exploit things perfectly. GPT 5.6 got a 78.5% on exploit bench. It also probably means we need a new benchmark for this if it aced it. So Astra will be able to be used for code review and patching, but it will refuse to create proof of concept exploits because it's too good at it.
The model was reviewed by the US government, but I wouldn't read too much into that. That's not really a rigorous review. They kind of tell the US government, hey, this is the model. If we don't hear from you in 30 days, we're going to put it out. And that's kind of what happened. It's more like a safety stamp. They didn't see anything that was obviously concerning, so they didn't stop it. OpenAI president Greg Brockman, however, said he believes Astra qualifies as AGI, artificial general intelligence, which is the term for something that is generally smarter than humans. There's not really a widely agreed benchmark on this. It's a I know it when I see it kind of thing, and Brockman thinks he sees it. He said, I quote, I leave it up to the reader to decide for themselves if this qualifies for them. I think we're there.
OpenAI says Astra's agentic capabilities can use a computer like a human. They gave a lot of examples, video game development, electrical engineering, bunch of scientific uses, fully interacting with and manipulating Excel spreadsheets. Those are just some of the things it can do with minimal human oversight. It's supposed to be very good at complex multi-step tasks. A lot of times agents, if you give them too many steps, they start to run out and then they come back like, what was I doing or they do something wrong? This is supposed to be really good at those kinds of things. And OpenAI says it's better at staying on task and respecting boundaries, which as we have recently learned is very important, and understanding user intent. A lot of benchmarks flying around out there. It scored 92.7% on Screenspot Pro. That is a test of understanding how to interact with software interfaces. That's a big jump. That's why I point this one out. 92.7% compared to GPT 5.6, which was 76.9%, which is why they're saying it's so good at doing things on a computer like you would. It should be noted that independent benchmark provider Epic AI rated it the most capable out of 267 models across 50 benchmarks, while artificial analysis, which focuses more on knowledge, coding and text comprehension, said it was even with GPT 5.6 and slightly behind Fable 5.1 in its estimation. Astra was built on OpenAI's largest ever training run. If you wondered what that Stargate site in Texas was good for, well, those 100,000 GPUs were used to train Astra.
And pricing for Astra, if you're using the API, is price matched. We've got Walmart style price matching in the world of AI. They'll be charging pretty much what entropic charges for Mythos or Fable 5.1, $10 per million input tokens, $50 per million output tokens.