Which model rewards caching the most?
Every provider discounts cached input tokens differently — some by 90%+, some far less. Enter the part of your prompt that repeats across requests and see exactly what caching is worth, model by model.
Figures cover only the repeated-context tokens above — not your full bill. See the calculator to project total monthly spend.
Why the discount varies so much by provider
Prompt caching works the same way everywhere in principle — a provider stores the computed state for a prompt prefix so a later request that starts with the same prefix skips recomputing it — but the discount each provider actually passes on to you is a business decision, not a law of physics, and it varies widely. Some providers price a cache hit near-zero; others give a smaller cut. That gap changes which model is the better choice once your traffic has a large repeated prefix, even if the two models’ standard input prices look similar.
What counts as "repeated context"
Only the part of a prompt that’s byte-for-byte identical across requests benefits — a fixed system prompt, a set of few-shot examples, a tool/function schema, or a retrieved document set that doesn’t change between calls. The variable part (the user’s actual message) never benefits from caching, since it’s different every time. Enter only the repeated portion above; mixing in the variable part will overstate your savings.
How the numbers here are calculated
Discount % is 1 − (cached rate ÷ standard input rate), using each model’s own live rates from our pricing table. Without cache assumes every request pays the full input rate for the repeated tokens; with cache blends the hit rate you set — at a 60% hit rate, 60% of repeated-context tokens bill at the cached rate and 40% at the standard rate, matching how a real cache miss (cold start, expired TTL, or a session outside the cache window) actually bills.
What this doesn't model
Some providers charge more than the standard rate the first time a prefix is cached (a "cache write"), before later reads become cheap — this calculator only models the discounted read side, since write costs are typically a small fraction of total spend once a prefix is reused many times. It also doesn’t include your output tokens or the variable part of each prompt; use the calculator for a full monthly bill projection.