GPT-6 Astra completed all 48 levels of the game ‘I’m Not a Robot’
OpenAI’s new GPT-6 Astra model passed all 48 levels of the browser game “I’m Not a Robot.” This is a CAPTCHA challenge created by Neal Agarwal, where each level requires proving you’re human anew: first just check the “I’m not a robot” box, then recognize images, read distorted text, and tackle much more exotic tasks. OpenAI engineer Sharif Shameem demonstrated the completion—the model relies on computer vision and computer control.
The blog author decided not just to retell the news, but to test the game himself. His own automation stopped at the second level. The first level — the familiar “I am not a robot” checkbox — the script passed by simply clicking the coordinates. But already at the second level, where you need to select from a 4×4 grid all squares with a stop sign, the general script could not solve which squares to mark. He had to look at the screen with his own eyes and manually specify the coordinates of four squares.
By the third level, the requirements change three times: “can you click?”, “do you see and understand?”, “can you read?”. At the third level, you need to read wavy distorted symbols — the author admits that even a human finds this difficult. There are 48 levels in total, and each essentially sets a new definition of “humanness”. Passing all 48, according to the author, means breaking through 48 different barriers separating machine from human.
The significance of the event is that captcha as a bot barrier is working increasingly worse. Previously, automation often hit a captcha wall; now this assumption is becoming outdated. But the author emphasizes the difference between “can pass” and “allowed to pass”: the terms of service of most services prohibit automated access.
For website protection, captchas alone are no longer enough, the author argues. He proposes alternatives: rate limiting, continuous behavior monitoring, verification via payment or identity confirmation, and a ‘slow down, not stop’ tactic to make bot attacks unprofitable. All single-point assessment mechanisms become unreliable over time – this has already happened with images, text, and audio.
For ordinary users, the main takeaway is that captchas have become so annoying precisely because machines have learned to pass them. Verification criteria are tightening, and users increasingly have to prove they are not robots. The author’s own comparison is telling: his automation stopped at the second level, while GPT-6 Astra passed all 48 – this gap shows the scale of current changes.
Primary source: note.com ↗