On September 7, 2026, Sharif Shameem, an engineer from OpenAI’s Labs team, posted a 4-minute video on the social network X. The video highlighted the capabilities of GPT-6 Astra, the latest model developed by OpenAI, which had been officially introduced just four days earlier on September 3. In the video, Astra successfully completed all 48 levels of a game called I'm Not a Robot, created by Neal Agarwal, a developer known for his work on the neal.fun website. The game, launched in September 2025, became popular online, with over 100,000 views on solution videos and speedrunners completing it in just a few minutes. The game challenges players to solve increasingly complex CAPTCHAs—security measures designed to distinguish humans from automated bots—by chaining them together. GPT-6 Astra is the latest and most advanced model in OpenAI's GPT series, following GPT-5 and its various iterations. Earlier this year, OpenAI mentioned that Astra might have crossed a high-risk threshold in its internal evaluation framework, particularly regarding its ability to perform tasks that could be considered hacking-related. The company describes Astra as its most capable model for "computer use," including tasks like browsing the internet, moving a cursor, filling out forms, and solving on-screen issues. Solving CAPTCHAs is one of the tasks Astra was specifically trained to handle, as these are a common part of online interactions. OpenAI rolled out GPT-6 Astra in stages starting on September 3, 2026. Initially, access was limited to a small group of organizations, followed by gradual availability to ChatGPT Plus, Pro, Business, and Enterprise users, as well as developers through OpenAI’s API. Some of the model’s most sensitive features, especially those related to cybersecurity, remain restricted to controlled access. While completing the I'm Not a Robot game is just one example, it signals a larger trend: AI systems are becoming increasingly adept at solving CAPTCHAs, which were once considered a reliable way to block bots. This trend has been developing for years. As early as 2025, it was already possible for AI models like ChatGPT to solve some traditional CAPTCHAs. Since then, AI models have improved rapidly, prompting companies to create more sophisticated verification systems. Tech giants like Google and Cloudflare are now exploring alternatives to the simple checkbox, such as analyzing user behavior or requiring more complex actions from users to verify their humanity. These efforts aim to make it more difficult for AI to bypass the verification process.