Yesterday it was announced that OpenAI's newest model, GPT-6 Astra, has saturated the ARC-AGI-3 challenge. It scored and impressive 99.9% with a Provider Adapter Harness. To put this into perspective, the best baseline score on ARC-AGI-3 was 30.2%, achieved by Anthropic Claude Opus 5 in July 2026.

Beating the AGI Challenge

I recently tried to build my own agent harness to tackle the ARC-AGI-3 challenge. My agent did well for the first demo game. However, the performance was nowhere near the human capability.

ARC-AGI-3 is designed to force agents to come up with novel strategies and understand a changing gaming environment by building world models. Up until yesterday, human's could beat these challenges much better than AI agents. This challenge requires the player to learn from mistakes in real time, form hypotheses, and use learnings from previous levels to build an understanding of game mechanics.

The fact that Astra was able to saturate this challenge with such a big jump compared to previous attempts should alarm the AI community. This is a sign that we have achieved the capability of creating AGI. OpenAI executives proudly state that Astra is the first model in the age of AGI.

AGI is Here

Personally, I think we can call this AGI or at least the first version of it. However, the creators of the ARC-AGI challenge disagree. In the blog post written by the ARC-AGI team, Greg Kamradt states that "saturating the benchmark would not represent proof of achieving AGI". Their main reasoning being that, though the ARC-AGI-3 challenge pushes agents to synthesize casual world models, it is far from an agent navigating the complexities of the real world.

In my opinion, if a model is able to synthesize even a casual world model abstractly, then moving to different environments is merely a question of providing the model different modes of input and output with some training to optimize for those. Essentially, I believe we will reach real-world AGI soon. We now have the proper engine through model systems like Astra.

What's Missing?

Like I've stated in past articles, the bridge between where we are now and full autonomous agents that can replace AI researchers is the memory substrate. I stand strongly by this. The key to advanced cognitive systems is intricate and directed organization and retrieval of memory.

AI models have gotten so good at reasoning that we don't have much more improvement to make in terms of their raw reasoning power.

If we want to get to the domain of Artificial Expertise, not only do agents need to be able to perform test-time learning (similar to what Astra has shown with its performance) but they must also be able to effectively store and retrieve learnings and context from years of consolidated test-time learning. That is what makes industry experts.

The Nuance of Domain Expertise

I've been a software engineer for around 8 years. I use AI regularly in development and research work today. When I first joined the industry, I believed that the best engineers were the ones that were technical wizzes; the engineers that could code a graph search algorithm with the lowest time complexity or create complex recursion loops. However, especially with the advent of AI reasoning, I've come to realize that what makes a great engineer is much more than raw intelligence. It is a combination of that and years of nuanced context assembly.

When you sit in a software engineering meeting full of PMs, different engineering teams, and cross-functional partners, something more is happening than simply designing a technical solution. If you ask a seasoned engineer "can we build a solution to this customer's problem?", the engineer is weighing business goals, technical tradeoffs, resources, timelines, and much more. A lot more is happening in their head than "can we code this?".

All of these tradeoffs and abstract calculations happen in the engineers head and within the context of conversations almost instantaneously. They pull from years of experience and combine it with knowledge from the current situation. That is an industry expert. His expertise strictly narrowed to operating cross-functional engineering decisions. He probably couldn't design a musical symphony and that's okay.

The Future is Expert Systems

After this advancement, I'm excited for the next step of intelligent systems. I think the future of AGI is Expert Systems: AI agents that can consolidate, organize and retrieve information and abstractions across days, months, and years of experience while keeping storage size and cost low.

Event with the advancements of Astra, our knowledge work is safe from AI takeover… for now.