In September 2026, OpenAI released its latest flagship model, GPT-6 Astra, with company president Greg Brockman proclaiming that "humanity has entered the AGI era." The U.S.-based ARC Prize Foundation, which operates the AGI benchmark ARC-AGI-3, publicly refuted the claim, stating Astra scored just 62.7% under standard conditions, far short of the 99.9% OpenAI had touted. On September 3, OpenAI officially launched GPT-6 Astra, describing it as "the most intelligent and most aligned model to date," with benchmark data showing capabilities surpassing predecessors across mathematics, programming, cybersecurity, professional work, and computer operation, plus native closed-loop agent capability requiring no external frameworks. The ARC Prize Foundation acknowledged Astra completed 96% of tasks on ARC-AGI-3 using fewer operations on average than humans, achieving "human-level trial-and-error efficiency," but issued an unequivocal no on whether AGI has arrived. The dispute centers on measurement conditions: under the foundation's neutral "standard testing conditions," Astra scored 62.7%; under "individually optimized conditions" with OpenAI's proprietary tooling, the same model's score soared to 99.9%—a 37-percentage-point gap. The difference concerns whether the model can retain its intermediate reasoning; under standard conditions, memory is wiped after every operation, while OpenAI's dedicated environment allows the full intermediate reasoning chain to carry into the next step