GPT-6 Astra sparks benchmark debate with human-beating ARC-AGI-3 efficiency

openai's gpt-6 astra receives mixed benchmark results, with epoch ai ranking it first among 267 models and artificial analysis rating it equal to its predecessor. however, gpt-6 astra achieves 62.7% on arc-agi-3, outperforming humans for the first time, and prompting arc prize chief francois chollet to revise his agi forecast. chollet considers the progress '2x faster' than expected. gpt-6 astra also shows improvements in efficiency, requiring fewer compute steps and tokens than its predecessors.