If you’re using Cursor with a custom OpenAI-compatible relay (to save costs or route through a gateway), you might not actually be getting the model you think you are. I’ve seen a few relays claiming to serve Sonnet 3.5 but secretly downgrading to Haiku when traffic spikes.
I built Relaylens to test for exactly this. It sends a signature-probe to any compatible endpoint and gives you a pass/fail grade in about a minute.
Works with any OpenAI or Anthropic compatible URL. Figured this crowd would appreciate the sanity check 