Astra’s proof claim fuels debate over artificial general intelligence and whether AI has surpassed human cognitive abilities

Different definitions of AGI

OpenAI president Greg Brockman has suggested that its advanced model, GPT-6 Astra, could mark the arrival of artificial general intelligence, or AGI. OpenAI defines AGI around an AI system outperforming humans at most economically valuable work, including coding, organisational tasks and project management.

Google DeepMind founder Demis Hassabis has offered a wider measure, saying AGI should manage “pretty much any cognitive task that humans can do”. The competing definitions reflect an ongoing debate about when AGI will arrive and how it can be recognised.

At a closed September briefing about Astra, Brockman said the industry had expected a clear moment of recognition, but that events had not unfolded that way. Anthropic chief executive Dario Amodei later called for the industry to slow development; Sam Altman and Elon Musk signalled agreement within hours.

A different way to work

Astra’s most noticeable change may be how users interact with it. Earlier models required detailed directions at each step; Astra is designed for longer tasks, with users setting a goal and boundaries while the model plans the steps.

The article reports that OpenAI estimates the average human scored 48% on the same unfamiliar games. ARC Prize, which runs the benchmark, reported Astra’s best result through its shared harness was 62.7%.

OpenAI’s 99.9% result came through a separate harness it built for Astra. That setup retained the model’s reasoning between requests, allowing it to reuse previous work rather than begin again each time.

Tests leave questions open

Independent evaluators disagree about how much Astra advances on GPT-5.6 Sol. Artificial Analysis did not rate the model as radically ahead of its predecessor. On September 8, 2026, OpenAI published what it described as a solution to the Navier-Stokes problem, one of seven Millennium Prize problems. Astra helped formalise and verify the work, a process that took a further 17 hours.

Warnings about AI risks

Amodei’s call to slow development came alongside concerns about the consequences of AI failures, which could disrupt key systems and cause damage possibly worth hundreds of billions of dollars. The report’s example involved a human choosing the targets and a model supplying the labour.