GPT-6 Astra Scored 99.9% on ARC-AGI-3: Did OpenAI Achieve AGI?
When GPT-6 Astra posted a 99.9% score on ARC-AGI-3, a benchmark built to test how well AI handles new and unfamiliar problems, social media declared that artificial general intelligence had arrived. The real story is more interesting. The same model also scored 62.7% under the benchmark’s standard rules, and the difference between those two numbers says a lot about how AI progress is measured, reported, and sometimes oversold.
What ARC-AGI-3 Actually Tests
ARC-AGI-3 is the third generation of the ARC-AGI benchmark series, run by the independent ARC Prize organization. Earlier versions focused on visual puzzles. This one is interactive: an AI agent is dropped into an unfamiliar game and has to work out what is going on.
Imagine opening a game with no tutorial, no rules, and nothing telling you what the goal is. You would start pressing buttons, watch what changes, form a guess about the rules, and slowly work out what the game wants from you. That is essentially the challenge an AI faces in ARC-AGI-3. The agent is not told the goal. It has to:
- Explore by trying actions and observing the results
- Infer the rules and the objective
- Plan an efficient way to win
The benchmark also measures efficiency, not just success. Its core metric compares how many actions the AI needs against how many a human needs, so wandering aimlessly is penalized. This design makes it hard to solve by memorizing training data, because the environments are meant to be new.
Until recently, even the most advanced AI models struggled with it. Early in 2026, the best frontier models scored in the low single digits or below, while humans solved the tasks reliably. That context is why Astra’s result drew so much attention.
Why GPT-6 Astra Has Two Different Scores
On September 3, 2026, ARC Prize published its analysis of GPT-6 Astra. It reported two results for the same model, tested under two different setups, which it calls harnesses. A harness is the scaffolding around a model: how it receives information, what it can remember between steps, and how long conversations are handled.
62.7%
Standard harness
A minimal, provider-neutral setup that every model is tested on. The model can carry forward notes it chooses to keep, but it does not keep everything. Cost for the run: about $26,000.
99.9%
Provider adapter harness
A setup that uses OpenAI’s own context-management features. It preserves the model’s reasoning state between requests and compresses long runs so earlier work can be reused. Cost for the run: about $19,000.
Scores as reported in launch coverage of ARC Prize’s results. Models were not all tested under the same harness, so the bars are not a perfectly like-for-like comparison. That is exactly the point of this article.
Both figures are real, and ARC Prize described both as state-of-the-art. The point is that they answer different questions. The standard result asks how models compare under the same minimal rules. The adapter result asks how well a model performs when its maker’s tooling is allowed to help.
The Memory Analogy That Explains the Gap
The difference between the two setups comes down to memory, and a simple analogy makes it concrete.
Standard harness: a few notes
You are solving a puzzle and are only allowed to keep a few notes from your last attempt. You might write “blue square bad,” but you lose the details of what you tried, what worked, and what did not. On the next attempt, you have to rediscover a lot.
Provider adapter: full recall
Now imagine you remember everything you tried. Each attempt builds on the last instead of starting over, and the puzzle gets easier as you go.
In technical terms, the standard setup requires the model to decide which information is worth saving as it moves through the test. The provider adapter preserves the model’s internal reasoning state between requests and compacts long conversations so earlier work can be reused.
Try actions in an unknown game
Save what was learned
Build on earlier attempts
Use fewer actions
This matters because ARC Prize has said a future AGI should be able to solve ARC-AGI-3 under standard conditions. Astra’s 62.7% in that setting is a strong result, but it is not the near-perfect score that made headlines.
What ARC Prize did praise
None of this makes the result unimpressive. ARC Prize reported that, in the provider adapter setup, Astra used fewer actions than the median human baseline on about 96% of the levels it completed, and averaged roughly 51.7% fewer actions per level. By the benchmark’s measure of action efficiency, that matches and surpasses human parity. ARC Prize also noted that Astra often turned unfamiliar environments into compact symbolic models of how they worked, which is a meaningful sign of progress toward generalization.
Why This Is Still Not AGI
ARC Prize described Astra’s result as meaningful progress toward generalization, but it did not call it AGI. There are several reasons the label does not fit yet.
1. It is a controlled environment
ARC-AGI-3 tasks have clear rules and a defined solution, even if the AI has to discover them. The environments are closed and deterministic, unlike the open-ended, messy real world.
2. Scaffolding played a role
The headline score used extra context-management tooling. A model that performs far better with its maker’s tools than without them is showing something about the whole system, not just the model alone.
3. The skills are not new
Planning, using tools, and working through several steps are already possible with today’s AI systems. Doing them very well is impressive, but it is not the same as being generally intelligent.
Reporting on the launch also noted that OpenAI’s own leadership described the release as the start of an “AGI era.” That is a marketing-flavored claim, while ARC Prize, the group that actually runs the test, took a more careful position: a near-saturated benchmark is a milestone, not proof of AGI.
What Would Actually Count as AGI?
There is no single agreed definition, which is one reason the debate keeps repeating. Two useful ways to think about it:
| Dimension | Question it asks | Where Astra’s ARC-AGI-3 result fits |
|---|---|---|
| Depth | How well can the AI do a task? | Very strong on this benchmark’s game-style tasks |
| Breadth | How many different kinds of tasks can it handle? | Not demonstrated by one benchmark |
| Autonomy | How independently can it operate? | Operates alone inside the game, but within a controlled setup |
Researchers at Google DeepMind have framed AGI along similar lines, looking at how capable a system is, how general it is, and how independently it acts. A simpler everyday definition is that an AGI should be able to learn any skill a human can, and do it with roughly the same efficiency. A single benchmark, however well designed, cannot test all of that.
A 3-Question Checklist for AI Headlines
The next time a post claims that AGI has arrived, run it through these questions before you share it:
Is it a narrow benchmark or a broad range of real work?
Did the system get extra tools, memory, or context?
Does the ability hold up outside the benchmark?
What This Means for Everyday Users and Businesses
- Read benchmark claims carefully. A headline number without its testing conditions can be misleading, even when it is technically true.
- Tooling matters as much as the model. The Astra results show that how a model remembers and manages context can change outcomes dramatically. That is useful to know when you choose or build AI tools.
- Agents are improving fast. Whatever you call it, AI is getting much better at acting on its own, exploring, and planning over many steps. That has real consequences for software, research, and business workflows.
- Expect this debate to return. At the current pace of progress, the AGI conversation will likely resurface within months, possibly weeks.
Frequently Asked Questions
Did GPT-6 Astra achieve AGI?
No. Astra posted a record result on ARC-AGI-3, but the benchmark is a controlled environment, the 99.9% score used the model maker’s own scaffolding, and ARC Prize itself did not call the result AGI.
Why are there two scores, 62.7% and 99.9%?
They come from two different testing setups. 62.7% was achieved under ARC Prize’s standard, provider-neutral harness. 99.9% was achieved with a provider adapter that preserves the model’s reasoning state between requests and compacts long runs.
Is a 99.9% score on ARC-AGI-3 still impressive?
Yes. Under both setups the result set a new state of the art, and in the adapter setup Astra used fewer actions than the median human on roughly 96% of levels. It is a real advance, just not proof of general intelligence.
What does “harness” mean?
A harness is the scaffolding around an AI model during a test: how it receives information, what it can remember between steps, and which tools or context-management features it can use. Different harnesses can produce very different scores for the same model.
What would convince experts that AGI has arrived?
There is no single agreed test. Most definitions involve performing well across a wide range of tasks, handling genuinely new problems, and operating independently, ideally with human-level learning efficiency.
Sources and Further Reading
- ARC Prize: OpenAI’s GPT-6 Astra on ARC-AGI-3 (primary source for the scores, harness definitions, and action-efficiency figures)
- Research & Development World: GPT-6 Astra scores 62.7% on interactive reasoning benchmark, near-perfect with custom adapter
- WION: Astra’s 99.9 per cent came from the harness, not the model
This article is for informational purposes. Benchmark results, model names, and leaderboard rules change quickly, so check ARC Prize’s website for the latest figures. Scores for other models are taken from launch coverage and may not have been measured under identical conditions.