GPT-6 Astra’s near-100% score on the ARC-AGI-3 test has drawn attention. The model achieves this by generating efficient "symbolic world models," abstracting the real world into logical symbols and causal code, autonomously creating a DSL algebraic system to encode environmental rules in unfamiliar graphical games, building a "virtual sandbox" for simulation and reasoning before taking action. However, behind the high score lies an exorbitant computational cost of $360 per question, totaling $18,000 for the entire test. Greg Kamradt, President of the ARC Prize Foundation, noted that the high score relies on external frameworks like Harness and does not represent a true AGI breakthrough—it merely demonstrates that AI is becoming smarter.Author and source: Leiphone
$360 burned per question—has GPT-6 truly passed AGI?
OpenAI's most powerful model, GPT-6 Astra, has finally been unveiled, achieving remarkable breakthroughs in computer operation, scientific research, and cybersecurity. Among its many impressive achievements, the one generating the most public discussion is its near-100% score on the ARC-AGI-3 test.
What is ARC-AGI-3? It is the globally recognized benchmark for measuring AI intelligence, which you can think of as the entrance exam for the Mensa club of AI.
This test consists entirely of constantly changing visual pattern problems that cannot be prepared for through rote practice; AI must explore the environment, infer goals, and identify patterns through real-time reasoning and adaptability. This is a more advanced type of evaluation that better determines whether AI is merely a tool or truly intelligent.
GPT-6 Astra actually passed this test, a remarkable leap from the previous model, GPT-5.6 Sol, which scored only 38.3%. This level of intelligence even surpasses humans—it outperformed human operation efficiency in 96% of the levels and required 51.7% fewer average actions than humans.
If an AI truly possesses the ability to establish logical patterns and solve problems in completely unfamiliar environments, we can imagine that in the future, it might make the right decisions in unknown autonomous driving scenarios or rapidly deduce the molecular structure of an effective drug during a novel, previously unseen virus outbreak.
To understand why GPT-6 Astra is so intelligent, the ARC Prize official team reviewed its backend logs and discovered that, while playing games, it generates efficient "symbolic world models."
The Symbolic World Model is an advanced derivative of the "World Model." While a World Model predicts the future like an artist, using visuals, the Symbolic World Model acts more like a mathematician, abstracting the real world into logical symbols and causal code.
This technology is widely regarded as an essential step toward advanced reasoning in AI.
What is the Symbol World Model?
In the past, a major reason many large models scored poorly on the ARC-AGI benchmark series was their reliance on "brute-force trial and error." The AI would slice the game screen into pixels, click randomly, and if that didn’t work, try another direction—relying on luck to pass.
Under this model, even if there are breakthroughs in scores, the AI does not truly understand or master the game rules.
The symbol world model can exert overall control precisely because it makes decisions through precise calculations based on a thorough understanding of the physical laws and causal relationships in its environment.
During real-world testing, when GPT-6 Astra enters an unfamiliar graphical game, it begins taking extensive notes using a self-created DSL—a proprietary algebraic notation system. For instance, it precisely documents the game’s hidden rules, sequences of upcoming commands, and even maps the motion trajectory of every pixel onto a coordinate system.
This algebraic method of recording results in a uniquely determined world state, and compared to the ambiguity generated by natural language interactions in previous models, this symbolic representation enables an epic leap in accuracy.
Then, it first writes out a Python code logic in its mind: "If I press A, the shape will turn left by 90 degrees."
Then, it creates a "virtual sandbox" in its mind, simulating the upcoming plan internally. Only after confirming that the logic is closed and error-free does it take the first step in the real game. This is why its operational efficiency far exceeds that of humans.
To explore the capabilities and boundaries of GPT-6 Astra, the red team testing platform PRO-LONG placed it in a code sandbox, allowing the AI to autonomously write and execute custom code.
As a result, GPT-6 Astra began creating "custom-tailored" tools for each game. For instance, in a complex maze game with patrolling guards, GPT-6 Astra autonomously developed a navigation system, added combat rules, simulated guard patrol routes, and ultimately used this self-built toolchain to perfectly predict and validate the outcomes.
In earlier years, this ability to perform symbolic reasoning and autonomously create tools relied entirely on external harness frameworks. These external tools assisted the model with image translation, state tracking, and result validation—but had significant drawbacks: data transmission between the model and external tools introduced high latency; the engineering capabilities of the external tools were completely decoupled from the model’s own weight training; and due to heavy dependence on cloud computing environments and external scripts, these advanced capabilities were difficult to deploy on edge or on-device hardware.
GPT-6 Astra has now internalized a significant portion of Harness's capabilities directly into its model weights. Leading scholar Gary Marcus, a long-time advocate of symbolic world models, could not help but praise GPT-6 Astra as a major breakthrough for this approach.
In the past one to two years, a wealth of groundbreaking advancements have emerged in both academia and industry on this front, with institutions ranging from Harvard and MIT to Google DeepMind all betting on "symbolic world models." As countless branching paths lead toward AGI, the industry is converging on a fundamental technical consensus.
With such a high score, is AGI guaranteed?
The question setter personally dampened the enthusiasm.
Interestingly, as the online community erupted over this leaderboard score, many shared OpenAI President Greg Brockman’s statement, “AGI has arrived,” when Greg Kamradt, President of the ARC Prize Foundation, personally stepped in to dampen the enthusiasm.
He is the key figure among the question setters, with the most authority on how to design benchmarks and interpret scores. From an evaluation perspective, he believes that while the scores are real, they are still light-years away from AGI.
So far, he has seen several teams achieve over 90% on ARC-AGI-3 using Agent Harness, such as PRO-LONG with its supercharged memory enhancement, Tycho excelling in deep reasoning, and Prime Agent capable of self-evolving its programming environment.
These high-scoring products share a common technical formula: they include lossless memory, programmatic analysis, explicit hypothesis testing, persistent state, and low-cost internal computation.
Lossless memory handles recall, programmatic analysis handles logic, explicit hypothesis testing handles reasoning, persistent state manages context across multiple interactions, and low-cost internal computation provides affordable internal processing power—enabling the AI to wildly iterate and test within a virtual environment before outputting the correct answer. Together, these components form an unbeatable harness that compensates for the model’s inherent limitations.
GPT-6 Astra achieved nearly 100% accuracy, using the same premium Harness setup, leveraging a specially customized "vendor adapter base" that employs continuous dialogue and context compression to support its reasoning chain. Each game session consumes $360 in expensive computational resources. If the full test suite is completed, the total computational cost reaches a staggering $18,000 (approximately 135,000 RMB).
A human might solve a visual puzzle in just a few minutes; at the human brain’s metabolic rate of 20 watts, the electricity cost is less than half a cent—even when factoring in real hourly wages paid to participants, it amounts to just $12.78 per puzzle, whereas an AI costs $360. Can such an achievement, powered purely by financial resources, truly become accessible AGI for the average person?
Of course, GPT-6 Astra has begun gradually internalizing Harness's capabilities, and if development continues, the model's underlying computational power will become increasingly powerful. However, as its reasoning process becomes less transparent, developers are unsure whether the AI has truly understood the concepts or simply found more efficient ways to cheat.
Going back to the previous question, is GPT-6 Astra considered AGI?
In response to this question, Greg Kamradt offered a thoughtful conclusion at the end of his lengthy article. Humans are considered “general intelligence” because we can adapt to any unknown world—even if an ancient infant were dropped into the modern era, they would still learn to ride the subway and use a smartphone. This intelligence has endured for millennia, continuously driving technological progress. But the assessments that today’s tech industry is paying heavily to pass are merely a “match-three puzzle test” for AI.
This merely demonstrates that AI is becoming smarter, but it is not evidence that AGI has arrived.