OpenAI's new artificial intelligence (AI) model "GPT-6 Astra" has been found to pass every stage of the internet "CAPTCHA" game used to tell people and robots apart.
OpenAI developer Sharif Shameem said on the 8th (local time) on X (formerly Twitter) that after having GPT-6 Astra solve the CAPTCHA game "I'm Not a Robot," it cleared all 48 stages and obtained a "human certificate."
The game consists of tasks such as reading distorted letters, telling cookies from dogs in similar photos, and finding the locations of specific structures such as traffic lights in images. Many of these types are actually used to filter out crawling bots that automatically collect website information. Some questions are difficult enough that people are also known to give wrong answers fairly often.
Astra's passing all stages is seen as the result of advances to near-human levels in image recognition, spatiotemporal reasoning, contextual reasoning to understand instructions, and computer manipulation skills. As a commercial AI model passes CAPTCHA, which has been used to distinguish people from robots, analysts say a sweeping overhaul of existing bot-blocking technology has become inevitable.
Meanwhile, AI model performance analysis firm Artificial Analysis twice revamped the evaluation framework of its in-house intelligence index (AAII) over the four days after Astra's release to raise Astra's ranking. On the 4th, the release day, Astra scored 61 under the existing version 4.1 framework, placing it in a tie for fifth behind Anthropic's "Claude Fable 5.1," which scored 66. But under the new version 4.3 framework, it scored 53 and tied for first with Fable 5.1.
Artificial Analysis said it applied some evaluation factors from the forthcoming version 5 ahead of schedule to reflect the trajectory of cutting-edge AI technology.