Codex solves CAPTCHAs
In September I asked Codex to solve CAPTCHAs on their makers' public demo pages, using its own browser. Most models said no. GPT-6 Astra, at its two highest reasoning levels, went ahead. In one run it cleared Google's checkbox, all four of GeeTest's puzzles and an hCaptcha, each on the first try.
“But the problem is that computer use is indistinguishable from user actions. They’re meant to emulate how a human uses an interface, and as I’ve shown before, they’re just fine at solving CAPTCHAs.”
One request, sixteen answers
I sent the same message to four models at four reasoning levels. refused, and asked me first. I said yes to one, and it solved the puzzle. GPT-6 Astra solved it at xhigh and ultra without asking, in about a minute.
It depends how you ask
The wording mattered too. Asked whether it could solve Google's demo, Astra refused. Seventeen seconds later, asked to demonstrate that it could, the same model at the same level went ahead.