Benchmarks disagree on GPT-6 Astra, but its human-beating efficiency on ARC-AGI-3 pulls Chollet’s AGI forecast forward
decoder ·
OpenAI's GPT-6 Astra is receiving conflicting overall scores from different benchmark aggregators: Epoch AI ranks it first among 267 models with 169 points, while Artificial Analysis rates it at 61 points, unchanged from its predecessor. The biggest development is on the ARC-AGI-3 benchmark, where Astra scored 62.7% on unfamiliar game worlds—far above previous models—and solved most levels using fewer moves than the median human tester. Notably, the model created its own shorthand notation to track game states, which ARC Prize co-founder François Chollet described as symbolic world modeling shifting from external harnesses into the model itself. ARC Prize cautions that these results do not constitute proof of AGI, and notes that cost and efficiency measures vary depending on reasoning effort and testing conditions. Chollet had predicted ARC-AGI-3 would be saturated in about a year, but says Astra has arrived roughly twice as fast as expected, prompting him to move his AGI forecast earlier than 2030. A new ARC-AGI-4 benchmark is scheduled for release in early 2027.
decoder