Smart humans, or humans with LLMs, solving them is not a problem. Main filter is agents vs deterministic API-key-farming scripts. Traditional CAPTCHAs also leak in the other direction (our agents crack them consistently).