Prompt caching savings calculator
Related Tools
Count tokens with o200k-style tokenization, then compare estimated LLM API cost across 30+ models. Pairs with the per-token cost calculator; all in-browser.
Enter input and output token counts; compare price across 30+ models. Deep-links from the token counter; cache- and reasoning-aware.
Side-by-side matrix: context, API $/M, capabilities; shareable ?models= links. Pairs with token counter and cost calculator.
Hub page listing every AI utility: token counter, cost calculators, RAG, caching, and programmatic model pages.
Users × requests/day × tokens with ×2/×5 stress and 12‑month projection for one reference model.
Rough planning estimate for image width/height → vision tokens.
Duration-based illustrative cost bands for speech and TTS-style workloads.
Float32 GB estimate plus embedding API math for ingest and monthly queries.
TPM and RPM headroom against illustrative dev/production tiers.
Break-even requests after a one-time fine-tune fee using your $/M assumptions.
Prompt + history size vs model context with 70% / 90% markers.
Editorial log of llm-pricing.json updates; pairs with /feed/pricing.xml.
Documents × chunks × top‑K × daily queries × output tokens — illustrative monthly LLM spend.
About Prompt caching savings calculator
This calculator isolates one piece of LLM economics: repeated system or tool prompts that qualify for provider prompt caching. You choose a model that has a cached-input price in our dataset, set how many tokens repeat each request, what fraction of those tokens are served from cache, and how much fresh output you generate. We compare hypothetical all-fresh input pricing against a split of fresh versus cached input using published dollars per million tokens. Cache-write fees, minimum billable units, and TTL rules are not modeled — use this as directional guidance only.
Key Features
- Models filtered to those with cached_input_per_mtok in llm-pricing.json
- Cache hit rate slider 0–100%
- Side-by-side per-request cost without vs with caching
- Percent savings on prompt-input spend
Use Cases
- Long system prompts in support bots
- RAG templates reused across users
- Teaching finance how cache hits change gross margin
Pro Tips
- If your hit rate is unknown, try 0.5 and 0.9 as a band.
- Pair with the token counter to estimate real prompt sizes.
Frequently Asked Questions
Why is my provider missing?
We only list models where our dataset includes a cached input rate; others may still support caching with different metering.