OpenAI just released GPT-6 Astra.
But one number immediately stands out: 99.9% on ARC-AGI-3.
Considering that ARC benchmarks were created specifically to study the gap between current AI systems and general intelligence, that number sounds almost absurd.
So the obvious question is:
Did we just get AGI?
Not quite.
But what Astra actually did may be more interesting than the headline.
ARC-AGI-3 is not a normal benchmark
Most AI benchmarks give a model a question and ask it to produce an answer.
ARC-AGI-3 works differently. The model enters an unfamiliar interactive environment without being explicitly told all of the rules.
To succeed, it has to:
- explore the environment,
- figure out how things work,
- infer what goal it should pursue,
- build some representation of the world,
- plan actions,
- and adjust when those actions don't work.
That distinction matters. An AI that recalls something from training is impressive.
An AI that encounters a system it hasn't seen, works out its mechanics, and then develops a strategy is testing something much closer to adaptation.
And Astra performed remarkably well.
ARC Prize reported that Astra sometimes transformed unfamiliar environments into compact symbolic representations—essentially creating its own shorthand for objects, rules and plans while solving the task. It also used fewer actions than the median tested human on 96% of evaluated levels.
That's a significant result.
But there is an important detail hiding behind the 99.9%.
The 99.9% score needs context
Astra was evaluated using two different setups.
Using ARC Prize's Standard harness, which gives different AI systems a common provider-neutral interface, Astra's best result was: 62.7%.
Using OpenAI's Provider Adapter, which preserves Astra's internal reasoning state between interactions and performs context compaction, the score increased to: 99.9%.
That doesn't make the second score invalid.It tells us something different.
The Standard result asks:
How capable is the model under a common interface?
The Provider Adapter result asks:
How capable can the entire Astra system become when the model is allowed to use the infrastructure designed specifically for it?
For increasingly agentic AI systems, that distinction may become very important.
Intelligence may no longer be determined by the model alone.
Memory, context management, tools and the surrounding inference system can fundamentally change what that model is capable of doing.
So is Astra AGI?
Even ARC Prize says no.
Their definition of AGI is extremely ambitious: roughly, a system capable of acquiring skills humans can acquire with comparable efficiency.
ARC-AGI-3 measures only part of that.
Its environments are deliberately unfamiliar, but they are still bounded, deterministic and closed-ended.
Real life isn't.
Real-world intelligence involves incomplete information, changing objectives, social reasoning, long time horizons, contradictory evidence, unpredictable environments and problems where nobody knows the correct answer beforehand.
ARC Prize therefore explicitly states that saturating ARC-AGI-3 should not be interpreted as proof that AGI has been achieved.
And I think that's the most important part of this release.
The AGI conversation may be changing
For years, the AGI discussion often sounded like:
“Can an AI solve problems it hasn't seen before?”
Astra suggests that frontier systems are becoming dramatically better at exactly that.
The next question may be harder:
How broadly does that ability transfer outside carefully defined environments?
Can a system continue learning when goals are ambiguous?
Can it recognize when its understanding of the world is wrong?
Can it operate reliably for days rather than minutes?
Can it handle environments where the rules themselves change?
And can it do all of this without humans carefully designing the surrounding task?
Those questions can't be answered by one benchmark.
Not even one with “AGI” in its name.
But something important did happen
Calling Astra “AGI” would be premature. Ignoring its result would be equally mistaken.
A system that can enter unfamiliar environments, construct useful internal representations, discover rules through interaction, and efficiently pursue goals is displaying a form of adaptive intelligence that frontier models struggled with only a short time ago.
ARC-AGI-3 was designed to expose exactly that gap. Astra just made the gap considerably smaller.
So maybe the most interesting question isn't:
“Did OpenAI achieve AGI?”
It's this:
If benchmarks designed to measure the road toward AGI are already approaching saturation, what abilities will we need to test next?
Because eventually, the hardest test of intelligence may no longer be whether an AI can solve a puzzle we've designed.
It may be whether it can understand a world we haven't.
Astra may not be AGI, but it has definitely moved the conversation forward.
So I am curious:
If you had to define the one capability that separates an extremely capable AI from true AGI, what would it be?
I'd genuinely love to hear different perspectives on this.