GPT-6 Astra Nears Perfect Score on ARC-AGI-3, but ARC Prize President Warns It Is Not AGI

An official review by ARC Prize of Astra's backend logs found that the model builds an efficient symbolic world model as it plays. Instead of relying on brute-force trial and error by clicking pixels, Astra uses a self-created DSL, an algebraic symbol system, to take extensive notes on the environment, recording hidden rules, sequences of upcoming instructions and the coordinate trajectories of every pixel. According to the review, the symbolic notation produced uniquely determined world states and avoided ambiguity from natural-language interaction. The model then wrote code-like logic in its own mind, such as a rule that if it pressed A, the figure would turn 90 degrees left, and simulated the plan in a virtual sandbox before making the first move in the real game, explaining its high efficiency.

During red-team testing on the PRO-LONG platform, GPT-6 Astra was placed in a code sandbox and allowed to write and run its own code. In a maze game with patrolling guards, it created a navigation system, added combat rules and simulated guards' patrol routes, then used the self-made tool chain to predict and verify the outcome. Leiphone says such symbolic reasoning and self-directed tool building used to depend on an external harness, which caused latency, separated engineering from model training and made advanced abilities hard to deploy on edge hardware. GPT-6 Astra has internalized much of that harness capability into the model weights. The report describes symbolic-world-model proponent Gary Marcus as praising Astra as a major victory for that research direction.

At the same time, many people shared OpenAI president Greg Brockman's line that AGI is here. Greg Kamradt, president of ARC Prize Foundation, responded with a warning. He said the score is genuine but far from proof of AGI. Kamradt noted that several other agent systems have passed 90 percent on ARC-AGI-3 when paired with a harness, including PRO-LONG with strong memory, Tycho for deep reasoning and Prime Agent with a self-evolving programming environment. Such high-scoring systems share five components, he said: lossless memory, programmatic analysis, explicit hypothesis testing, persistent state and low-cost internal computation, with the last enabling a large number of trial-and-error attempts in a virtual environment before an answer is output.

According to Kamradt, Astra's near-perfect result also used a luxury harness. The system relied on a custom vendor adapter base, continuous dialogue and context compression, and consumed about 360 dollars' worth of compute per game; a complete run of the test would cost roughly 18,000 dollars. A human can solve one of these puzzles in minutes, and by Kamradt's estimate, the electricity cost of a 20-watt human brain is less than half a cent. Even after adding a real participant's hourly wage, a human solution costs 12.78 dollars per game.

Kamradt concluded that such a costly benchmark is only a graphical match-3 exam for AI. It proves that AI is becoming smarter, he wrote, but it is not evidence that AGI has arrived. Human beings are generally intelligent because they can adapt to an entirely unknown world; an infant from an ancient era, placed in the modern world, would still learn to ride the subway and use a mobile phone. That adaptability, he argued, is not measured by the test.