A model exploring a question runs many searches, most of which return little. Metering the searches would punish exactly the behaviour that makes an agent useful; metering the results does not. Fifty free results is enough to find out whether the corpus covers your field before any card is involved.
BioloGPT's endpoint, served over SSE. The free tier needs no key at all — you get 50 results before anything is metered.
Search across their biology corpus. The endpoint did not return a tool list to our probe, so the surface here is what the vendor documents rather than what we enumerated.
Nothing for the first 50 results. After that, an API key and $0.02 per returned result — searches themselves are unlimited, and you pay only for results that come back.
Paste a URL — add the endpoint to your client