Every gateway advertises caching. The detail that matters is the billing rule: on SwitchGate, a cache hit costs exactly zero credits — not a discount, not a reduced token rate. If we didn't call the provider, you didn't pay anyone.
Two caches, one rule
Exact-match caching handles the boring, enormous case: identical request bodies (same model, messages, temperature) within your configured TTL. Semantic caching is opt-in per key and matches paraphrases — "cancel my subscription" and "I want to cancel my plan" — using an embedding threshold you control. Both are scoped to your organization; there is no cross-tenant cache, full stop, because your prompts are not another customer's speedup.
Where 30–70% is real (and where it isn't)
The savings figure on our homepage is labeled a placeholder, but the shape is honest: support bots, documentation Q&A, and RAG frontends with repetitive query distributions routinely see hit rates north of 30%. Agentic coding and long-context analysis see almost none — every request is genuinely novel — and we'd rather tell you that than sell you a mirage. The ledger shows your actual hit rate per key from day one, so the question resolves itself with data.
Hits appear on the ledger as $0.000 line items with a cache: hit trace tag. Turning semantic matching on is one toggle per key; turning it off is the same toggle, and nothing about your integration changes either way.