The papers can make it free, they can just build a convenient search/retrieval API layer so Open AI and Anthropic don't pay 1-2 engineers a year at 1-2m to crawl and index that info for training.

The AI services have an option then to pay for this service, support a open service, or write their own crawler. I think if every open AI request didn't just do a web search but a more targeted arXiv search the results would be better.