[AI SPRINT] AGI Arrived on Saturday. Your Tomorrow Looks the Same.

I think of it this way: compare it to the engine in your car. Say it can do 140 miles an hour. That doesn't mean you'll ever drive 140. It doesn't mean your tires and brakes are rated for it. And it certainly doesn't mean the roads you drive on were built for it. Raising the engine’s top speed changes almost nothing about your commute.

That's where most companies sit right now.

What Astra actually is

Last Thursday, OpenAI announced GPT-6 “Astra.” The 99.9% benchmark score it achieved is impressive, but came from a custom setup OpenAI built for its own model. Run the standard version of that same test, the one every other model takes, and Astra lands at 62.7%. That’s a real improvement over all other models—the next closest is Claude Opus 5 at 35.2%, and the prior flagship GPT-5.6 is at 7.8%.

Astra is better at operating software on its own, moving through a browser, working across applications, and finishing a multi-step task without someone watching each step. It does that work in roughly half the time the previous model needed. Astra is impressive, and all of this is what leads people to claim Astra is “AGI”.

Capability went up—but so did the bill. It’s about two and a half times more costly to run. For those organizations that enable it (it does need to be enabled), expect to get a lot of requests for the Premium $125 a month plan soon after.

So what is AGI?

AGI stands for “artificial general intelligence”, and nobody owns the definition. It generally means an AI with enough knowledge and judgment to reliably exceed what humans can do, and that description shifts by the month, by the group you ask, and by whatever a company happens to be selling that week.

Huang is a useful example, and I say that as someone who mostly agrees with him. In March, on Lex Fridman's podcast, the bar on the table was an AI that could build a billion-dollar company. Huang said "I think it's now. I think we've achieved AGI," then hedged that Fridman said a billion and didn't say forever. On his August earnings call, eight days before Astra shipped, he was more careful, saying that for many tasks we could say AGI was already achieved and that "all of those milestones are kind of senseless at this point." Then on Saturday he posted that AGI has arrived, with no definition attached, in the same four sentences as "400K GPUs coming online next."

All three statements are compatible with each other. That's what a term looks like when it has no fixed meaning. It stretches to fit the room.

It is worth noting who isn't celebrating too. After Huang posted, Jakub Pachocki, OpenAI's own chief scientist, published an essay arguing that no lab, his own included, has solved alignment and monitoring well enough to keep scaling at maximum speed much longer. He expects voluntary slowdowns to become common.

The company selling the chips announced arrival. The person building the model asked the industry to slow down.

Like Huang, I think we crossed some meaningful line earlier this year. Whether it was this model or the next one is a moot point. If Astra isn't it, something will claim the title by December.

So why does any of this matter?

It matters for your organization because the pace went up again, and every time it does, the same gap gets wider.

Many companies I speak with are barely scratching the surface of AI tools they've had for three years. After all, using AI to write an email is not using superhuman capability. It's using a very expensive spell checker. And what AI can do has increased materially four or five times since 2023, while the floor of use in companies has barely budged.

Here's what that gap actually costs. Two companies buy the same licenses on the same day. One hands them out and sends a training email. The other picks three workflows, assigns an owner to each, and rebuilds them around what the tool can now do. Twelve months later they are not the same company anymore, and no amount of catching up on licenses closes it. Every capability jump like Saturday's compounds that difference rather than resetting it.

Your customers, competitors, and investors don't care which model you licensed. They notice whether your service got faster, your quality got better, or your people got freed up to do work they couldn't do before.

What about it’s danger?

You've also probably heard that Anthropic’s head of AI safety quit, posting his reasoning online. He didn't say much directly about why he left, focusing instead on his belief that "the world is in peril. And not just from AI. Or bioweapons. But from a whole series of interconnected crises unfolding at this moment (the 'Polycrisis')."

Combined with Pachocki's letter on AI alignment, the recent reports of AI models acting outside their development environments, and the claims about AGI, another wave of concern about AI danger has begun.

That's a full subject for another newsletter, but I want to address it briefly here. My view is that it's valuable for these people to raise their concerns, and that we should debate them openly. Attention and debate are important steps toward making sure we use AI in a way that supports humanity, and this creates the attention that drives improvement.

That said, I don't believe humans face existential risk from AI itself. The more significant risk, and the one we already see, is people using AI to cause harm. That's been true of every technology ever created. We need to focus more on reasonably identifying and preventing malicious use than on the imagined existential risk of Skynet.

What to do about it this month

Here are three easy things you can do in your organization, right now:

1. Measure velocity instead of adoption. Count how many workflows in your company run differently today than they did 30 days ago. Not seats deployed, not licenses purchased. Workflows changed, each with a manager's name on it. If that number is zero, a more capable model changes nothing for you.

2. Pick the three workflows worth rebuilding. Not the ones AI can assist. The ones where a system that operates software end to end would let you deliver something you can't deliver today. That's a different list than the one your team would hand you.

3. Decide who owns the answer. Someone needs the job of turning capability into operating change, with protected time to do it. Without that person, every launch like this one is just news.

The pace of change will continue to increase. The organizations that come out ahead won't be the ones that gave everybody GPT-6. They'll be the ones that rebuilt their products, services, and operations around what it can do.

Saturday didn't change what you should do this week. It changed how much it costs to keep putting it off.

None of that is a tooling decision, which is why it doesn't get solved by a license. If you want to work through it alongside other leaders doing the same thing, join Leader Lab. You get the working sessions, the frameworks, and the tools we use with clients to make this concrete.

Also open to everyone this month: RSVP for the September 25th AI SPRINT meeting, where we'll recap the month and open the floor for discussion.

Trent Gillespie is CEO of Stellis AI and a keynote speaker helping business leaders understand and operationalize AI in their companies. He spent almost nine years leading global innovation efforts at Amazon before leaving to help other companies build the capabilities they need to compete. Book Trent to speak to your group or book a call to discuss using AI within your business.

Did someone forward this newsletter to you? If you're not already signed up, you can subscribe to AI SPRINT™ for free here.