0. Five-second answer: AI could solve elite problems yet still get lost in a simple game with no manual
AI has been strangely uneven.
It surpassed humans in chess long ago, and by 2025 an advanced Gemini Deep Think system earned an officially graded IMO gold-medal-level score of 35/42. Yet frontier models could still struggle when handed a simple-looking 2D environment with no instructions and asked to figure out what to do.[1]
ARC-AGI-3 was built for that gap. There are no natural-language rules or stated goals. An agent must interact, observe what changes, infer mechanics, identify useful goals, and plan.[2][3]
In September 2026, GPT-6 Astra reached 62.71% on ARC-AGI-3 Semi-Private with the Standard harness and a best verified 99.95% with OpenAI's Provider Adapter.[2][4]
That does not prove AGI. It is not 100%, and the 99.95% run used provider-specific context management.
The important change is narrower and more useful: AI became much better at learning inside an unfamiliar environment where the task itself is initially unclear.
For markets, that suggests a better role than “predict tomorrow's price.” An Astra-like agent is more interesting as a researcher that discovers candidate edges, tests them, and tries to break its own findings.
1. What is AGI? And 1, 2, 3 are benchmark generations, not AI levels
AGI means Artificial General Intelligence.
ARC Prize frames AGI as a system able to acquire skills humans can acquire, with human-like efficiency.[2]
ARC-AGI-1, 2, and 3 are not ranks such as “AGI Level 1, Level 2, Level 3.” They are generations of a benchmark.
- ARC-AGI-1: introduced in 2019; colored-grid abstraction puzzles where a solver infers a hidden transformation from a few examples.[5]
- ARC-AGI-2: the same basic format, but harder. Every included task was verified as solvable by at least two people within two attempts.[5][6]
- ARC-AGI-3: moves from static puzzles to interactive, unfamiliar environments that must be learned over many actions.[3]
For a child-friendly summary:
1 = visual puzzle
2 = much nastier visual puzzle
3 = a brand-new game after someone threw away the manual
2. What does an “unknown game” look like? A 2D world whose meaning appears only after you act
ARC-AGI-3 uses turn-based 2D grid environments. Game state can be represented on grids up to 64×64, and agents choose from available actions.[7]
The key distinction is that knowing which buttons exist does not tell you what winning means.
An invented example, not an official task, might begin like this:
■■□□□□
■●□□▲□
□□□■□□
No explanation.
You move right. The dot moves. You touch the triangle and its color changes. Somewhere else, something resembling a door opens.
A human starts forming hypotheses:
“Maybe the dot is me.”
“Maybe the triangle is a switch.”
“Maybe I need to change state before reaching the door.”
ARC-AGI-3 targets exactly this loop: exploration → modeling → goal acquisition → planning and execution.[2]
The benchmark is designed to be quick for humans to pick up, but that does not mean every child clears every game. The 2026 human study involved 458 participants. The environments were calibrated to be human-solvable, while individual success still varied.[8]
3. Why could AI be great at chess and hard mathematics but weak here?
Chess gives you the full rules. Mathematics tells you what must be proved.
Those problems can be incredibly difficult, but the problem frame is already supplied.
ARC-AGI-3 asks earlier questions:
“What counts as progress?”
“What does this object do?”
“What changed because of my last action?”
Humans do this constantly with unfamiliar devices, games, workplaces, and apps. We interact a few times and build a rough “probably works like this” model.
Frontier AI had a major weakness in forming and maintaining that kind of model from limited experience.
It could be brilliant at a formal theorem and then become confused by a toy-like interface.
That is why ARC-AGI-3 is useful: it exposes that intelligence is not one smooth scale.
4. GPT-5.6 Sol: a strong brain that was effectively throwing away its notebook
GPT-5.6 Sol Max scored 7.78% on the verified ARC-AGI-3 Semi-Private evaluation, despite 96.5% on ARC-AGI-1 and 92.5% on ARC-AGI-2.[9]
OpenAI investigated and found that part of the problem was not just the model but the harness—the software connecting model and game.[10]
In the official setup, hidden reasoning was discarded after actions, and older history was eventually truncated.
The human analogy is absurd:
“Stepping on red opens the door.”
→ take one step
→ tear up the note containing that discovery
→ next step: “What is this red thing?”
When OpenAI enabled retained reasoning and context compaction, Sol's score on the Public set rose from 13.3% to 38.3%.[10]
Do not directly rank 7.78 against 38.3: the former is Semi-Private and the latter is Public.
But the lesson is clear: capability depends not only on the model's raw reasoning, but also on how experience is preserved and compressed.
5. GPT-6 Astra: 62.71% and 99.95% are not the same kind of result
ARC Prize verified GPT-6 Astra on September 2, 2026.[4]
Its best Semi-Private results were:
| Evaluation condition | Best Astra score |
|---|---|
| Standard harness | 62.71% (Max) |
| Provider Adapter | 99.95% (High) |
The Standard harness is closer to a shared minimal interface: the model decides what visible notes to preserve.
The Provider Adapter additionally preserves provider-specific opaque reasoning state between requests and uses compaction for long contexts.[2][4]
So “Astra by itself scored 99.95% naked” is inaccurate. A better description is Astra plus OpenAI-designed context management scored 99.95%.
Even the Standard result is a huge jump from Sol Max's 7.78% on the same Semi-Private benchmark family.[4][9]
The human comparison is also easy to misstate. Astra Max under Provider Adapter scored 98.55%, and on the levels it completed it used fewer actions than the human baseline on 96.0% of levels, with 51.7% fewer actions per level on average.[2]
That does not mean “Astra is smarter than 96% of humans.” It is a level-by-level action-efficiency comparison against a median human baseline.
The real number is already dramatic enough without upgrading it into science fiction.
6. What was Astra actually doing? Building a tiny private physics textbook for each game
ARC Prize highlighted Astra's behavior, not just its score.
Instead of retaining every observation verbatim, Astra compressed information about:
- object locations,
- action effects,
- current state,
- unfinished plans,
- and the next useful operations.
It even created compact symbolic shorthand for game mechanics.[2]
Conceptually, it was producing rules such as:
red + right action → state B state B + blue target → door opens
This is a small, task-specific world model: a representation that predicts what the environment will do next.
That matters because the correct answer was not memorized in advance. The useful structure had to be extracted from interaction.
7. So is this AGI? ARC Prize explicitly says no
ARC Prize describes Astra as a noticeable step-function change in frontier capability, but it does not claim that Astra proves AGI.[2]
ARC-AGI-3 is still a bounded world:
- 2D grids,
- turn-based interaction,
- deterministic mechanics,
- closed-ended goals.
The real world contains changing rules, hidden variables, strategic opponents, unclear objectives, and environments where yesterday's pattern stops working today.
The competition's Grand Prize also requires 100%. A verified 99.95% is extremely close, but it is not 100%.[11]
The accurate takeaway is not “general intelligence is finished.” It is:
a capability that had been a major weakness—learning efficiently inside a novel interactive environment—just improved sharply.
8. Where could this matter? Jobs where nobody can write the full rulebook in advance
The most interesting applications are not more puzzle solving. They are systems where humans know what outcome they want but cannot fully specify the winning rule.
Examples include:
- Factory and logistics simulation: vary staffing, inventory, sequencing, and routing to discover conditions that reduce delay or cost.
- Software exploration: learn unfamiliar interfaces and identify reproducible bug conditions.
- Games and robotics: learn causal action effects with limited trials.
- Advertising, pricing, and operations: explore combinations of decisions and observe results.
- Research assistance: generate hypotheses, test them, reject failures, and update the model of the system.
The task changes from:
“Follow this rule.”
to:
“Find which rules actually matter.”
9. What about stocks and scalping? A candidate-edge discovery agent is more plausible than an oracle
Here, an edge means a small repeatable statistical advantage that still exists after costs.
An Astra-like approach would not start with a finished human rule such as “buy when RSI is below 30.”
Instead, it could inspect order-book states, trades, price, spread, volume, and time, then ask:
Which combinations of current conditions are followed by a small directional bias over the next seconds or minutes?
It might propose a hypothesis involving, for example, a sudden increase in bid depth, a burst of market buys, a narrow spread, and a particular time window.
That is only a hypothesis—not a profitable rule.
Markets differ fundamentally from ARC games. ARC mechanics are stable. Markets change as participants, news, liquidity, regulation, and regimes change.
If you simply tell an AI “find a profitable pattern,” it can mistake historical coincidence for a law of nature.
10. A credible trading system would make the AI attack its own discoveries
A backtest uses historical data to evaluate a trading rule.
The danger is multiple testing: try thousands or millions of variants and some will look spectacular purely by chance. This is the well-known problem of backtest overfitting.[12][13]
A serious “edge explorer” should therefore run something like this:
- Generate hypotheses. What current states might matter?
- Search on a development period. Produce many candidates.
- Reject on untouched data. Test on periods not used for discovery.
- Split by market regime. Try to break the rule in rising, falling, volatile, and quiet conditions.
- Include trading frictions. Fees, spread, and slippage must be deducted.
- Record how many ideas were tried. One winner after 10,000 attempts is not automatically a miracle.
- Walk forward in time. Refit only on past data, then test on the next period repeatedly.
- Paper trade first. Test in real time without real money.
- Define kill conditions. Stop when the estimated edge disappears.
The ideal role is not prophet.
It is one researcher who proposes 1,000 ideas plus one reviewer who kills 998 of them.
That is much closer to translating ARC-AGI-3-style exploration into a hostile financial environment.
11. Final conclusion: from “AI that answers” to “AI that discovers what the rules are”
The important part of ARC-AGI-3 is not just the flashy 99.95%.
It is the learning loop:
observe → act → fail → infer a rule → revise the hypothesis → move toward a goal.
Sol showed how much a powerful model can suffer when its experience is poorly preserved. Astra showed a major improvement in compressing experience into useful rules and executing with few interactions.
That suggests a broader way to use advanced agents.
Instead of saying:
“Here is the rule. Execute it.”
we can increasingly say:
“Here is the environment and the data. Find the rules that seem to matter, try to falsify them under different conditions, and bring back only what survives.”
There is no guarantee that this produces money in markets. Markets are far nastier than ARC.
But if we stop treating AI as a fortune-teller and start treating it as a machine for discovering and attacking hypotheses, Astra's ARC-AGI-3 result becomes a very practical kind of news.
Sources
- Google DeepMind — IMO 2025 gold-medal-level result (2025-07-21) deepmind.google
- ARC Prize — “OpenAI's GPT-6 Astra on ARC-AGI-3” (2026-09-03) arcprize.org
- ARC Prize — “Announcing ARC-AGI-3” (2026-03-25) arcprize.org
- ARC Prize Verified Results — GPT-6 Astra (verified 2026-09-02) arcprize.org
- ARC Prize — ARC-AGI-1 & ARC-AGI-2 Guide arcprize.org
- ARC Prize — “ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems” (2025-05-20) arcprize.org
- ARC-AGI-3 Docs — Game Schema docs.arcprize.org
- ARC Prize — “Measuring Human Performance on ARC-AGI-3” (2026-04-14) arcprize.org
- ARC Prize Verified Results — GPT-5.6 arcprize.org
- OpenAI — “How two settings tripled our ARC-AGI-3 scores” (2026-07-29) openai.com
- ARC Prize 2026 — ARC-AGI-3 Competition; Grand Prize requires 100% arcprize.org
- Bailey, Borwein, López de Prado & Zhu — “The Probability of Backtest Overfitting papers.ssrn.com
- Bailey & López de Prado — “How ‘backtest overfitting’ in finance leads to false discoveries” (2021) academic.oup.com
