On 3 September, OpenAI president Greg Brockman ended a press briefing with five words: "Welcome to the AGI era."

Three days later, Nvidia's CEO, Jensen Huang, posted: "AGI has arrived."

So a lot of people looked up what AGI means. Views of Wikipedia's "Artificial general intelligence" page went from 46,816 in August to 125,720 in September, almost three times as many.

Source: Wikimedia Pageviews (monthly user views)

We have been here before. In 2023, Microsoft researchers wrote that GPT-4 showed "sparks" of AGI. Since then, someone has said AGI is close, or already here, almost every year.

This keeps happening for a simple reason. There is no agreed definition of AGI, so each company uses the one that suits it. Once you know the main definitions, AGI headlines get much easier to judge.

AGI stands for artificial general intelligence. The important word is "general". Today's AI can be very good at one task, like writing code. AGI is supposed to be good at almost any task a person can do. (If AI, ML and LLMs still blur together, start here.)

"Almost any task" is too vague to test. So different groups wrote their own definitions. These are the five you will see most often.

1. AGI can do most jobs. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". Put simply, it does most paid jobs better than people do.

2. AGI learns new skills as fast as a person. ARC Prize, the group that runs the ARC-AGI tests, says AGI is a system that can learn any new skill as quickly, and with as little practice, as a human can.

3. AGI comes in levels. A 2023 Google DeepMind paper splits AGI into five levels: Emerging, Competent, Expert, Exceptional and Superhuman. It put ChatGPT at the lowest level, Emerging.

4. AGI can make money. Mustafa Suleyman, a DeepMind co-founder, suggested a test: give an AI $100,000 and see if it can turn it into $1 million. Microsoft's deal with OpenAI once reportedly tied AGI to about $100 billion in profits.

5. AGI is whatever feels like AGI. There is no test here, only "I'll know it when I see it". This is the one Brockman used. Looking back in a couple of years, he told reporters, "I think it might be about this model."

So when someone says "AGI is here", the first thing to ask is which of these five they mean.

For years, "self-driving" meant whatever the company selling the car wanted it to mean. So the car industry agreed on six levels, from 0 to 5. At Level 2, the car steers and brakes, but you must watch the road the whole time. At Level 4 and above, the car drives itself.

The levels gave buyers a way to check claims. Tesla calls its system "Full Self-Driving", but by Tesla's own description it is a Level 2 system, so the driver has to keep watching. In December 2025, California's DMV ruled that the name "Full Self-Driving Capability" was misleading. The name now ends in "(Supervised)".

The question that told you the most about any car was simple: who has to watch the road?

DeepMind's levels for AGI copy this idea on purpose.

You don't need to pick one definition. You need three questions.

1. Who benefits from the claim? OpenAI is preparing to list on the stock market, and hopes to be valued at more than $850 billion. Huang's post ended with "400K GPUs coming online next", and Nvidia sells GPUs. Days before the launch, OpenAI's own CEO, Sam Altman, called AGI "an irrelevant marketing term". None of this makes the claim false. It tells you how much weight to give it.

2. What test was run? The AGI talk came from one score on ARC-AGI-3. This test puts an AI into small puzzle games with no instructions. The AI has to work out the rules and the goal by playing. Humans can solve every game. It is a good test of one skill: learning the rules of a simple game. It tells you nothing about whether a model can run a sales team. And OpenAI's own test of real paid work, GDPval, which is the closest match to its own definition, was not on the launch scorecard.

3. What setup was the model given? Engineers call this the harness. It is everything around the model: what it can remember, which tools it can use, and how its work carries over from one step to the next. Give the same model a different setup and it can get a very different score. (We saw this with LLM-as-a-Judge: a score is only as good as the setup behind it.)

The third question matters most for Astra.

OpenAI published a score of 99.9% on ARC-AGI-3, next to 30.2% for Claude Opus 5 and 7.8% for its own GPT-5.6 Sol. A footnote says Astra ran on OpenAI's "responses API harness", which "changes two settings to better match real-world performance".

¹ “On ARC-AGI-3, GPT-6 Astra was run with our responses API harness, which changes two settings to better match real-world performance. The changes do not specifically target ARC-AGI-3.”

Source: OpenAI (the ARC-AGI-3 row and footnote 1 from the launch page)

ARC Prize then tested Astra itself. On its standard setup, the same one every model gets, Astra scored 62.7%. On a second setup that let the model keep its hidden reasoning between moves, it scored 99.9%. The 30.2% and 7.8% for the other two models were standard-setup scores.

So OpenAI's table put Astra's best setup next to the other models' standard setup. With everyone on the same setup, Astra scored 62.7%. ARC Prize now reports both numbers, clearly labelled.

It depends on the definition. Here is how Astra does on each one:

And the people who built the test say no:

Toby Walsh of UNSW's AI Institute put it bluntly: "I'd eat my hat if we didn't find trivial things that an eight-year-old can do that Astra fails at."

Astra is a real step forward. On the same setup as everyone else, it scored about twice what Claude Opus 5 scored. OpenAI also says it has helped solve long-standing open problems in maths.

But calling it AGI is marketing. The only definition it clearly meets is the one with no test.

I think "is it AGI?" is the wrong question for anyone building products. The useful question is smaller: which tasks can I give this model and stop checking its work? With cars, the question was who has to watch the road. With AI, it is who has to check the output.

No press briefing will answer that for you. You answer it by testing the model on your own work, with your own examples. That is what evals are for.

So next time someone tells you AGI is here, ask three things: who benefits, what test, and what setup.

That's it for today,
Sid