OpenAI’s GPT-6 Astra recorded scores of 62.7% and 99.9% on ARC-AGI-3, depending on the test setup, leaving open whether it meets the economic-work standard in OpenAI’s AGI definition.
OpenAI’s announcement highlighted Astra’s 99.9% score. ARC Prize Foundation, which operates the benchmark, recorded 62.7% with its Standard harness at maximum reasoning effort and 99.9% with a Provider Adapter at high effort. ARC Prize listed costs of $26,098 and $18,817, respectively. Because the runs used different harnesses and reasoning settings, their results reflect separate setups.
The Provider Adapter carries an opaque reasoning state between requests and compresses long conversations, allowing Astra to reuse earlier work. ARC-AGI-3 contains 135 abstract environments with no instructions on how to proceed. Participants must experiment, infer each environment’s rules and decide which actions to take.
In the Provider Adapter run, ARC Prize said Astra used fewer actions than the median tested human on 96% of levels. That figure measures action efficiency within the benchmark; it does not establish performance across human cognitive abilities.
OpenAI’s Charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” It also says no firm timetable has been set for AGI. ARC Prize frames its benchmark around whether a system can learn human skills as efficiently as a person, a different measure from performance across economically valuable work.
OpenAI reported a 72.6% score for Astra on its OSWorld 2.0 computer-use evaluation, with each task taking roughly 40 minutes. GPT-5.6 Sol scored 65.7% at roughly 75 minutes per task. OpenAI’s Astra announcement did not include a GDPval score. That evaluation covers economically valuable tasks across 44 occupations and the nine sectors contributing most to U.S. gross domestic product, so the released results do not establish Astra’s performance across that broader category of work.
At a Sept. 3 briefing, OpenAI President Greg Brockman called Astra’s launch the start of an “AGI era,” Axios and VentureBeat reported. EconoTimes has previously examined how AI terminology and evaluation methods answer different questions. The broader performance claim remains unverified by the results reported so far.
