For this use case it matters a lot less if LLMs can solve it. As long as it costs you more to solve the captcha than it costs your adversary to serve it to you, it is still (some what) effective.