OpenAI's GPT-6 Astra on ARC-AGI-3 (arcprize.org)

OpenAI has announced its latest model, GPT-6 Astra, which has achieved groundbreaking scores on the ARC-AGI-3 benchmark, registering a 62.7% score using the Standard harness for $26K and an impressive 99.9% when utilizing the Provider Adapter harness for $19K. This advancement is significant as GPT-6 Astra not only meets but exceeds human performance benchmarks in action efficiency, using fewer actions than the median human tested across 96% of levels. The model's innovations include turning unfamiliar environments into compact symbolic world models and developing a domain-specific language for more effective planning and execution. ARC-AGI-3 serves as a critical benchmark for evaluating agentic intelligence in abstract environments, where agents must explore, infer goals, and build internal models without explicit instructions. The benchmark measures four components of agentic intelligence: exploration, modeling, goal-setting, and planning. Astra's capability to efficiently process and represent game mechanics demonstrates a leap in AI development, underscoring the model's potential to bridge the gap towards artificial general intelligence (AGI) while still acknowledging that ARC-AGI-3 does not equate to achieving AGI itself. OpenAI's results highlight not only the evolving capabilities of AI but also set the stage for future benchmarks that could further illuminate the path toward advanced generalized AI systems.