Interactive · Week 2 · Sections 3, 5

Budget your context like the resource it is.

Most teams underprice their LLM bill by an order of magnitude — and most of them could cut it by 60–80% by enabling prompt caching alone. Drag the component sliders, pick a model, set a volume, and see what your real cost looks like with and without caching. The checkboxes mark which components are static enough to cache.

Context components

Total input tokens 0

Cacheable 0 (0%)

Settings

Provider · model

Volume 100/min

Requests per minute. Monthly = volume × 60 × 24 × 30.

Cache hit rate 90%

Fraction of requests that find a warm cache. Production teams typically see 75–95% once the prefix is stable.

Per request · no cache

—

Per request · with cache

—

Monthly · no cache

—

at current volume

Monthly · with cache

—

Per-request cost breakdown · with cache

What's happening here

Prices are published list rates verified August 2026 — exact, not rounded — but vendors move them, so re-check before making a budget decision. Cache economics differ by vendor and are modeled separately: Anthropic bills cache hits at 10% of input with a 1.25× write premium; OpenAI bills hits at 50% of input with no write penalty; Google bills hits at roughly 10% of input with no write penalty, but also charges cache storage by the hour ($1.00–$4.50 per 1M tokens per hour) — which this calculator does not model, so Google's cached scenarios read cheaper here than they bill. DeepSeek's rates are approximate and unverified. Two things are deliberately excluded: long-context surcharges (Gemini 2.5 Pro doubles above 200k input tokens; the GPT-5.6 family switches to long-context rates above 272k) and Claude Sonnet 5's introductory $2/$10 pricing, which expires 31 August 2026 — the standard $3/$15 is used instead. Real bills will vary; this calculator is for relative reasoning, not invoice prediction. Source-of-truth for the production version lives as <PromptCostCalculator/> in the curriculum codebase.