Checking responses in production is also a good idea because it reveals all the edges that may not be exercised in an artificial test suite.

There is still some performance cost for checking a response against a schema though, so a nice compromise is just to check a small % of outgoing responses — the vast majority of requests stay as fast as possible, but given a reasonable traffic load, it's still enough to eventually reveal any places the schema doesn't match reality. (This is an approach we use at Stripe for checking our OpenAPI specification.)