Prompt Caching Cost Calculator: How Much Do You Actually Save?

It depends on your cache hit rate, not just whether you reuse a prompt. On Claude, a cache write costs 1.25x the normal input rate (5-minute TTL) or 2x (1-hour TTL), so a single cache write followed by zero reads costs more than not caching at all. The math works out in your favor once your hit rate clears about 22% on the 5-minute TTL or 53% on the 1-hour TTL, regardless of how large the cached prefix is. Below that line you are paying a premium for a discount you rarely collect. On GPT-5.4, caching has no write premium, so it never costs more than leaving it off.

The most common cause is TTL churn: the default cache entry expires after 5 minutes of inactivity, and every expiration forces the next request to pay the full cache-write premium again. Bursty traffic with gaps longer than 5 minutes between calls can end up paying the write premium repeatedly with few or no matching reads in between, which is more expensive than sending the prompt uncached. Switching to the 1-hour TTL reduces how often you re-pay the write premium, at a higher per-write cost, so it only helps if your gaps are longer than 5 minutes but shorter than an hour.

A cache write happens the first time a given prefix is stored: it costs 1.25x the standard input rate on the default 5-minute TTL, or 2x on the 1-hour TTL. A cache read happens on every subsequent call that hits the same stored prefix before it expires: it costs 0.1x the standard input rate, a 90% discount. These two multipliers are the same across the whole Claude lineup, Haiku 4.5 through Fable 5, so the break-even hit rate is a property of the TTL you choose, not the model.

No. OpenAI caches automatically once a prompt prefix is roughly 1,024 tokens or longer, and a cache miss is billed at the same rate as a normal, uncached request. There is no separate write premium the way there is on Claude. A cache hit is billed at 10% of the standard input rate. Because a miss never costs more than not caching, GPT-5.4 caching has no realistic downside, only upside on the requests that happen to hit.

Check the token usage fields on the API response rather than assuming caching is working. On Claude, look at cache_read_input_tokens and cache_creation_input_tokens in the usage object; if cache_read_input_tokens stays at zero across repeated, near-identical requests, something in your prompt (a timestamp, an unsorted JSON object, a rotating request ID) is silently breaking the prefix match. On OpenAI, the cached_tokens field inside prompt_tokens_details reports the same thing. Reported bugs where caching silently stops working, sometimes inflating costs 10 to 20 times for a session, are usually traced back to exactly this kind of prefix drift.

On Claude, the minimum cacheable prefix is model dependent: 4,096 tokens on Opus and Haiku tiers, 2,048 tokens on Fable 5, and 1,024 tokens on some Sonnet-generation models. A shorter prefix silently will not cache at all, no error, it just shows zero cache-read tokens. On GPT-5.4, OpenAI applies caching automatically above roughly 1,024 tokens with no configuration needed.

Use the 5-minute default when your requests against the same prefix are frequent, at least one every few minutes, since the write premium is lower (1.25x vs 2x) and you will re-pay it less often anyway. Use the 1-hour TTL when your traffic is bursty with gaps longer than 5 minutes but the same prefix still gets reused within an hour, for example a support bot that goes quiet overnight then gets a morning rush. The break-even hit rate is higher on the 1-hour TTL (about 53% vs 22%), so only switch if you are confident the longer window will actually raise your hit rate enough to clear it.

The discount on a cache hit is close, 90% off standard input pricing on both. The mechanics differ in two ways that matter for cost. First, OpenAI caching is automatic and requires no cache_control markers or opt-in; Claude requires you to mark which content blocks are cacheable. Second, and more importantly for your bill, OpenAI charges no premium on a cache miss, while Claude charges 1.25x to 2x. That asymmetry is why a low hit rate can make Claude caching a net loss while GPT-5.4 caching cannot go net negative.