Prompt Cache Savings Calculator

The break-even hit rate

Where a provider charges a premium to write an entry into the cache, a miss costs more than an ordinary uncached call. Caching only pays once hits are frequent enough to outweigh those writes. Setting cached cost equal to uncached cost and solving for the hit rate h gives:

Two consequences worth internalising. If writes are billed at the ordinary input rate, h* = 0 and any hit at all is free money. If writes carry a premium, there is a floor below which caching is a net cost, and a low-traffic service can quietly land under it.

The calculator marks that floor against your derived hit rate so the comparison is against your own traffic, not a vendor's best case.

A useful sanity check on the formula: Anthropic's pricing page states that caching pays back after one cache read on the 5-minute window, and after two reads on the 1-hour window. Feeding their published multipliers into the equation reproduces exactly that. A 1.25× write against a 0.1× read puts the floor at 21.7%, which two calls (one write, one read) clear at 50%. A 2× write puts it at 52.6%, which two calls miss and three calls clear at 66.7%.

Why traffic shape decides this, not prompt design

A cache entry expires after an idle window. If the average gap between two calls sharing a prefix is longer than that window, the entry is gone before the next call arrives, so every call is a miss no matter how carefully the prefix was constructed. This is why the same feature produces opposite outcomes for two teams running identical prompts.

The derived mode models this directly: it spreads your daily requests across your active window, compares the resulting gap to the TTL, and treats the cache as warm only when the gap fits inside it. When it does not, the hit rate collapses and the write premium applies to every call.

Methodology

• Cost without caching = requests × (prefix + dynamic) × input price
• Cost with caching = (writes × prefix × write price) + (hits × prefix × read price) + (requests × dynamic × input price)
• Average gap = active hours ÷ requests per day
• Derived hit rate = (requests − 1) ÷ requests when the gap fits inside the TTL, otherwise 0
• Monthly = daily × 30

Dynamic tokens are always billed at full input price. Output tokens are excluded entirely, because prompt caching discounts the reused input prefix and leaves completion billing untouched. The derived hit rate is a deliberately simple two-state model, warm or cold, which matches how a steady-traffic and a sparse-traffic workload actually behave; if you have a measured cached-token share from your provider, switch to manual and use it.

Prices are yours to supply. Any preset carries the source and the date it was checked. Vendor pricing and cache retention windows change, so confirm against the current official pricing page before you budget from these numbers.

When to use it, when to skip it

What Effloow added

Most caching calculators ask you for a hit rate, which is the number you least reliably know before you migrate. This one derives the hit rate from traffic shape and TTL, surfaces the break-even threshold explicitly, and warns when the shared prefix is too short for caching to engage at all.

The cost model was checked two ways before shipping. It reproduces the per-1,000-call arithmetic from our measured GPT-5.6 cache run, and its break-even formula independently reproduces the payback rule Anthropic publishes for its own 5-minute and 1-hour cache windows. Preset rates were read from each vendor's official pricing page on 2026-08-12 and carry that date in the tool; where a vendor does not publish a figure, such as Google's default cache TTL, the preset says so instead of guessing.

Measured background reading

• GPT-5.6 prompt caching: the 24-hour cache is gone — an API run measuring the write premium and the TTL ceiling.
• OpenAI's 24h prompt cache: we measured the real discount — when the discount appears, and when it does not.
• Token optimization for production LLM systems — the wider cost picture beyond caching.
• AI Token Estimator — measure a prefix before you paste its token count in here.