Instrument 05 / token budget

Know the monthly LLM bill before the first invoice.

Turn traffic, input and output tokens, current API prices, prompt caching, and retry overhead into one inspectable monthly estimate.

Model one month

Convert token assumptions into a budget.

Use current provider pricing and observed traffic when available. Nothing is submitted or stored; every calculation runs in this browser.

Requests per monthExpected model calls before retry overhead. calls Input tokens per requestSystem prompt, history, retrieved context, and user input. tok Output tokens per requestAverage generated tokens, including responses later rejected. tok Input price per 1M tokensUse the provider's current uncached input rate. $ Output price per 1M tokensGenerated-token prices are often different from input. $ Cached input shareThe input tokens that receive the provider's cache rate. % Cache discountThe percentage saved on eligible cached input tokens. % Retry overheadExtra calls caused by timeouts, validation failures, or regeneration. % Other cost per requestOptional retrieval, search, tool, gateway, or observability charge. $

Budget boundaries

Use current prices.
Provider rates and cache rules change; this calculator does not freeze them into a preset.

Count generated tokens.
Output can dominate spend even when prompts are long.

Measure retries.
Failed calls still consume tokens and attached services.

Transparent method

What this LLM API cost calculator computes.

Input spend is request volume multiplied by input tokens, retry overhead, and the effective input price. The effective price applies the stated cache discount only to the cached share.

Output spend uses output tokens and the separate generated-token price. Other usage repeats with retries because retrieval, search, gateways, or tools may run again.

Retry spend is the difference between the same workload with and without retry overhead. Cache savings compares the effective cached input price with the uncached input price.

This is a planning estimate, not a provider quote. It excludes taxes, committed-use discounts, tiered pricing, batch discounts, storage, fine-tuning, and provider-specific billing rules unless you represent them in the input prices or other per-request cost.

Token budget guide

Four questions behind a defensible LLM API estimate.

The model price is only one variable. Traffic, context size, generated length, caching eligibility, and failed calls decide what reaches the invoice.

Read the full AI evaluation guide → Need cache-read, cache-write, and break-even math? →

How do you calculate LLM API cost?

Multiply monthly input and output tokens by their respective prices per million tokens. Apply a cache discount only to eligible input tokens, then add retry overhead and any retrieval, search, gateway, or tool cost that repeats per request.

Why are input and output tokens priced separately?

Providers commonly charge different rates for tokens sent to the model and tokens generated by it. Output tokens often cost more, so a single blended token price can materially understate the bill.

How does prompt caching affect LLM cost?

Prompt caching can lower the price of repeated eligible input tokens. Use an observed hit rate where possible, and verify the provider's cache lifetime, minimum prefix, write charge, and invalidation rules.

How do retries affect token cost?

Retries repeat some or all of a request, adding input tokens, output tokens, and attached service charges. Estimate overhead from logs rather than assuming each user request creates exactly one billable call.