OpenAI's GPT-6 Astra Hits 99.9% on ARC-AGI-3 Benchmark
The model demonstrates a leap in agentic general intelligence by developing its own symbolic shorthand to solve novel environments.
OpenAI's GPT-6 Astra has achieved a near-perfect score on the ARC-AGI-3 benchmark, signaling a major shift in how large language models handle unfamiliar environments. The results suggest the model is moving beyond simple pattern recognition toward true symbolic reasoning.
According to the ARC Prize, GPT-6 Astra scored 99.9% on the ARC-AGI-3 benchmark when utilizing a Provider Adapter harness, a feat accomplished at a cost of $19,000. When tested using a Standard harness on the Semi-Private set, the model achieved a score of 62.7% with a total cost of $26,000. Beyond raw accuracy, the model demonstrated extreme action efficiency, outperforming the median human in the number of actions required to solve a level on 96% of the tests.
The Path to Agentic Intelligence
ARC-AGI-3 is specifically designed to measure agentic general intelligence by forcing models to navigate and solve problems in environments they have never encountered during training. Unlike traditional benchmarks that rely on static datasets, this test evaluates the ability to learn on the fly. This capacity for real-time adaptation is critical for moving AI from a passive information retriever to an active agent capable of solving open-ended problems.
Symbolic Reasoning and World Models
What distinguishes Astra's performance is not just the score, but the method of problem-solving. The ARC Prize reports that the model exhibited the ability to represent complex game mechanics as a set of logical rules. Most notably, Astra developed its own domain-specific language shorthand to create compact symbolic world models of these unfamiliar environments. This allows the model to abstract the rules of a system and apply them logically rather than relying on probabilistic guesses based on previous training data.
Implications for AGI
These results represent a significant leap in the quest for Artificial General Intelligence (AGI). By surpassing human efficiency in specific agentic tasks and inventing its own internal shorthand for reasoning, GPT-6 Astra demonstrates a capacity for autonomous adaptation. The ability to synthesize new logical frameworks in real-time suggests that LLMs are becoming capable of handling the "novelty gap" that has historically separated AI from human-level flexibility.
Future Outlook
As OpenAI continues to refine these agentic capabilities, the industry will be watching to see if this symbolic reasoning translates to real-world software engineering or scientific discovery. While the Provider Adapter harness yielded near-perfect results, the gap between that and the Standard harness performance remains a key area for technical analysis. It remains to be seen if these efficiencies can be scaled down to reduce the high financial cost of running such complex reasoning chains.