Well, That Didn't Take Long

A day after we argued that the AI stack is bending around continuity, Astra showed up with a benchmark result that makes the whole idea considerably harder to ignore.

There is a very specific kind of comedy that happens when you put an idea into public view and reality immediately decides to make it look less speculative.

Not prove it. Not validate some grand theory. Just walk into the room, drop a giant data point on the table, and leave. That's basically what happened here.

On August 30, 2026, in Everyone Is Yelling AGI. I Think They're Missing the Interesting Part, we were talking about how the AGI label itself may be less interesting than the machinery gathering around the model: memory, tools, browser access, computer use, long-running tasks, multi-agent coordination, persistent context, and all the plumbing that turns a model into something that can actually do work. The broader point was that capability can start compounding at the system level, not merely because the model weights change, but because the entire machine around the model gets better at preserving experience and using it.

Then, on September 2, 2026, we pushed that idea further in The AI Stack Is Starting to Bend.

That piece was basically about one thing: continuity.

The user doesn't want to explain themselves again. The agent shouldn't have to rediscover the work. The next model shouldn't have to reread the room. And increasingly, the machinery underneath all of this may be moving toward preserving and carrying forward state rather than continuously reconstructing everything from scratch.

The conclusion was deliberately cautious. Maybe memory, persistent agents, KV-state transfer, recurrent architectures and inference hardware are just unrelated engineering optimizations that happen to rhyme. Or maybe, a few years from now, we look back and realize they were pieces of one larger transition: from AI as a series of transactions to AI as a continuing cognitive process.

Cool.

We put the little marker down. And then ARC Prize published the Astra results.

Well. That didn't take long.

62.7 versus 99.9

The number everybody is going to talk about is 99.9%.

OpenAI's GPT-6 Astra scored 99.9% on the ARC-AGI-3 Semi-Private benchmark using what ARC Prize calls the Provider Adapter harness.

That is an absurd score.

Astra also scored 62.7% under ARC's standard provider-neutral harness, which was already state of the art.

Naturally, the internet is going to spend the next several days doing what the internet does.

"AGI IS HERE."

"No, it isn't."

"Yes, it is."

"Benchmark contamination."

"Benchmark saturation."

"Move the goalposts."

"François Chollet has entered the chat."

I'm much more interested in what happened between 62.7 and 99.9. Because the biggest difference between those two evaluation environments is not simply "more intelligence." It's how the system handles state.

ARC's standard harness gives the model the information needed to solve the environment and lets it decide what to preserve in visible notes.

The Provider Adapter harness does something more interesting:

It preserves opaque reasoning state between requests and uses compaction for longer interactions, allowing the model to reuse prior work instead of repeatedly rebuilding it.

There's that word again. State.

No, this doesn't mean "persistent state adds 37 points of intelligence"

Before somebody screen-grabs the previous paragraph and turns it into something stupid, no.

We can't simply subtract 62.7 from 99.9 and declare that continuity makes an AI 37.2% smarter. These are different harnesses with different context-management behavior. The reasoning settings are different in the peak runs. There are multiple variables involved. ARC Prize is very explicit about why it reports both.

The standard harness asks how models perform through a common provider-neutral interface.

The Provider Adapter asks a different question: How well does the model work when it gets to use the context-management machinery its own provider designed for it?

Those are both useful questions. But for the thing we've been discussing, the second one is almost comically relevant. Because one of the strongest frontier-model results we have just seen comes from an environment where the system is allowed to carry its internal work forward. Not just the words. Not just a transcript. The reasoning state.

And the efficiency difference is just as interesting as the raw score.

Across the game/reasoning pairs solved by both setups, ARC reports that the Provider Adapter runs were about 3.66 times faster and used 49% fewer tokens overall. That doesn't look like a minor UX feature. That looks like architecture starting to matter.

The model didn't just remember. It built a world.

The other part of ARC Prize's analysis may be even more interesting.

Astra apparently got very good at taking unfamiliar environments and reducing them into compact symbolic models. It tracked objects, coordinates, rules and unfinished plans. It created its own compact algebraic shorthand to represent what was happening. ARC describes it as a domain-specific notation the model generated on the fly so it could keep track of state and plan actions efficiently.

That matters because ARC-AGI-3 isn't simply asking the model to answer questions. The environments require exploration, modeling, goal-setting, planning and course correction. The system has to encounter something unfamiliar, figure out the rules, build a working representation of the environment, then use that representation to act.

That is a very different shape from:

prompt → answer

It looks much more like:

observe → model → preserve → act → update → continue

Which is suspiciously close to the shape we just spent an entire article talking about. Again, not proof. Just annoyingly on-theme.

Everybody is yelling AGI again

And of course the funniest part is that this arrives right on top of the AGI debate.

ARC Prize itself is not claiming Astra is AGI. They explicitly say that saturating ARC-AGI-3 would not constitute proof of AGI, because the benchmark is still bounded, deterministic and far less complex than the open-ended real world. Good. That's the responsible position. And it's also why I remain much less interested in fighting over the label.

Back in Everyone Is Yelling AGI, we were talking about how AGI has become one of those words that almost collapses under the weight of everyone's definition. Researcher definition. Economist definition. Engineer definition. Twitter definition.

My own definition has always been much more practical: Can I give the system serious work and have it reliably get the work done?

That's why the system around the model has been more interesting to me than the ceremony around declaring AGI. And Astra is giving us another example of why.

OpenAI has already designated Astra its first model at the Critical cybersecurity capability level, meaning that with the proper tools and access it can find previously unknown vulnerabilities and develop exploit strategies across hardened systems without a human guiding every step. It hit 100% on ExploitBench and, in OpenAI's own evaluations, discovered previously unknown vulnerabilities and chained them into working exploits.

That is obviously significant.

But even there, the interesting part is increasingly not simply the IQ of the model. It's what happens when you give the model time, tools, state and a continuing environment in which to operate.

The harness is becoming part of the intelligence

This is the thing I think people are going to have to get used to.

We have spent the last several years treating "the model" as if it were the product. Model A beats Model B. Model B gets a bigger context window. Model C gets a better benchmark score. But the system is starting to complicate that picture.

A frontier model inside a minimal stateless interface is one thing.

The same frontier model inside a system that preserves its reasoning state, compacts old context, gives it tools, lets it operate software, lets it maintain notes and lets it continue working across long horizons is something else.

At some point, the harness becomes part of the capability. And that is exactly what we were talking about when we said the wall is increasingly becoming plumbing. The model can be brilliant and still fail because it lost context. Because it had no access. Because it had no tools. Because it couldn't preserve what it learned three steps ago. Because every new invocation effectively woke up in a hotel room with no idea how it got there.

Fix enough of those things and suddenly the exact same model starts looking dramatically more capable. The intelligence was there. The continuity wasn't.

This is where the stack bends

That phrase felt slightly metaphorical when we wrote The AI Stack Is Starting to Bend. Now it feels a little less so.

At the user layer, we increasingly want systems that understand us over time. At the agent layer, we want work to persist across tasks and sessions.

At the state layer, we are trying to preserve already-computed context instead of rebuilding it. At the model layer, recurrent approaches allow internal state to be processed repeatedly. At the hardware layer, inference architectures are being designed around increasingly persistent workloads.

And now we have a benchmark result where a model's strongest performance appears when the evaluation setup explicitly lets it preserve opaque reasoning state and reuse prior work. Different technologies. Different mechanisms. Different layers. Same pressure. Don't make the process start over.

That is the thing I keep seeing.

And then the efficiency result sneaks up on you

The 99.9% number is going to get the headline.

I actually think the efficiency result deserves almost as much attention.

ARC found that Astra, using the Provider Adapter harness, used fewer actions than the median human baseline on 96% of levels and averaged 51.7% fewer actions per level.

That matters because one of the longstanding differences between humans and AI systems has been experience efficiency. A person can walk into an unfamiliar situation, poke around a little, develop a rough mental model and then act. AI systems have often needed far more attempts, far more exploration and far more brute force.

ARC's own analysis notes that once Astra appears to understand the mechanics of an environment, its behavior becomes much more human-like in terms of how efficiently it acts. That sounds suspiciously like maintaining a useful world model. Which brings us right back to state.

Again.

Apparently state is now following me around.

So where does this leave the "I seent it" thing?

Exactly where it was supposed to be.

As a joke. Not a prophecy. Nobody predicted Astra from a Möbius strip.

Nobody gets to claim ownership of recursion because they drew a loop on a whiteboard. And this benchmark doesn't prove that the whole AI stack is inevitably converging on some persistent cognitive superstructure.

That would be ridiculous.

But we did put a public marker down saying something very specific:

Maybe the important transition isn't simply toward bigger and smarter models. Maybe it's toward systems where useful cognitive state persists long enough to compound.

Then, on September 3, 2026, one of the most dramatic benchmark results of the year shows up in a configuration built around preserving reasoning state between requests. I am absolutely not saying "called it." I'm just looking at the little marker. Then looking at the benchmark. Then looking back at the marker. And moving it about three feet closer to the corkboard.

Very quietly. I seent it.

Sources & Further Reading

ARC Prize — OpenAI’s GPT-6 Astra on ARC-AGI-3 Primary source for the 62.7% Standard Harness score, 99.9% Provider Adapter score, preservation of opaque reasoning state between requests, 3.66× faster runs, 49% fewer tokens, symbolic world-model behavior, and action-efficiency results. (ARC Prize)https://arcprize.org/blog/astra

OpenAI — Path to Astra: Critical Capabilities and Frontier Safeguards OpenAI’s official disclosure that Astra meets its Critical cybersecurity capability threshold, including its exploit-development capabilities and safeguards. (OpenAI)https://openai.com/index/path-to-astra/

OpenAI — Pacing Model Development in an Era of Cyber-Critical Capabilities Background on the security controls, development pacing, monitoring, and containment measures surrounding increasingly capable cyber models. https://openai.com/index/pacing-model-development-cyber-capabilities/

ARC Prize — ARC-AGI-3 Background on the benchmark, including its focus on exploration, world modeling, goal-setting, planning, and execution. https://arcprize.org/arc-agi/3

Wes Roth — “AGI IS HERE” The video that surfaced the new Astra results and kicked off this follow-up discussion. https://www.youtube.com/watch?v=3rSFxrKAb3o

The AI Stack Is Starting to Bend The earlier article arguing that the emerging AI stack is increasingly reorganizing around continuity and persistent cognitive state. (Rich Washburn)https://www.richwashburn.com/post/the-ai-stack-is-starting-to-bend

Everyone Is Yelling AGI. I Think They’re Missing the Interesting Part. The earlier piece arguing that the more important development may be the growing system around the model—memory, tools, agents, persistence, and long-running work—rather than the AGI label itself. (Rich Washburn)https://www.richwashburn.com/post/everyone-is-yelling-agi-i-think-they-re-missing-the-interesting-part

---

Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.