Question 1

How are LLM API costs calculated?

Accepted Answer

LLM APIs charge separately for input tokens (the prompt + context) and output tokens (the generated response). Cost = (Input tokens / 1,000,000) × Input rate + (Output tokens / 1,000,000) × Output rate. Example using illustrative rates of $3/M input and $15/M output: a call with 1,000 input tokens and 500 output tokens costs (1,000/1,000,000 × $3) + (500/1,000,000 × $15) = $0.003 + $0.0075 = $0.0105/call. At 100,000 calls/month that is $1,050/month. The per-million-token rates are user inputs in this calculator — always confirm the current number on the provider's published pricing page, because LLM prices change often.

Question 2

Why are output tokens so much more expensive than input tokens?

Accepted Answer

Output tokens are consistently several times more expensive than input tokens across all major providers. The reason is architectural: generating each output token requires a full forward pass through the model, one token at a time, while input tokens are processed in parallel because attention is computed over all of them simultaneously. The practical implication is that applications that generate long responses (creative writing, detailed analysis, verbose chat) are fundamentally more expensive than applications with short responses (classification, extraction, short answers). The fastest cost lever in most apps is capping max output tokens and prompting for terse answers — that is exactly the surprise this calculator is built to surface before you ship.

Question 3

What is a token and how many tokens is a typical message?

Accepted Answer

A token is roughly 4 characters or 0.75 words in English (it varies by language and model). Rough benchmarks: a 100-word user message is about 130 tokens; a 500-word system prompt is about 650 tokens; a 1,000-word document chunk is about 1,300 tokens; a 500-word generated response is about 650 tokens. To get exact counts for a specific string, use the provider's tokenizer — OpenAI publishes tiktoken and Anthropic ships a Claude tokenizer. Once you have realistic input and output token counts, drop them into this calculator with the prices you pay to see the real per-request and monthly economics.

Question 4

Can I trust the example prices, and how do I keep them current?

Accepted Answer

No — treat every example price as a dated illustration, not a live quote. This calculator deliberately does not hardcode any provider's pricing in its math, and it does not fetch live prices. The example presets on the page are labeled with the month they reflect and exist only to save you typing; the moment a provider changes a rate, the preset is stale. The right workflow is: open the provider's published pricing page (Anthropic, OpenAI, Google, or whoever you use), read today's per-million-token input and output rates, and type those numbers in. Because the formula is timeless and only the inputs carry a date, the result is always correct as long as your inputs are.

Question 5

What is prompt caching and how do I model it here?

Accepted Answer

Prompt caching (offered by Anthropic, OpenAI, Google, and others) stores the processed state of a repeated prompt prefix so subsequent calls with the same prefix pay a steep discount on the cached portion. Cache-read rates are commonly a small fraction of the standard input rate. To model it, put the repeated, cached portion of your prompt in the cached-input-tokens field and the cache-read price in the cache-read field; put only the new, uncached tokens in the standard input field. If you leave the cache-read price blank, the calculator bills cached tokens at the standard input rate — it never assumes a discount you did not state. For a dedicated savings model, see the Prompt Caching Savings Calculator.

Question 6

How do I estimate the cost of a RAG (retrieval-augmented generation) app?

Accepted Answer

RAG applications have two token-cost components. First, retrieval: vector search returns k document chunks that are injected into context. If each chunk is 500 tokens and you retrieve 5 chunks, that is 2,500 additional input tokens per query. Second, generation: the full prompt (system + user message + retrieved chunks) is billed at the input rate, and the answer is billed at the output rate. Model it here by setting input tokens to (system + user + retrieved-chunk tokens) and output tokens to your average answer length, then enter your monthly query volume. If your system prompt and chunk templates are stable, move the repeated portion into the cached-input field to capture caching savings.

Question 7

What is the effective price per 1M tokens, and why show it?

Accepted Answer

The effective price per 1M tokens is the total cost of a request spread across every token it touched — input, output, and cached — expressed as a single blended unit price. It exists because raw input and output rates are hard to compare across models when their input/output split differs. A model with cheap input but expensive output can cost more than a model with balanced rates, depending on your workload. The effective price collapses your specific token mix into one number you can line up against another model's effective price for the same workload, which is the honest way to compare providers.

Question 8

Does this include batch API discounts or fine-tuning surcharges?

Accepted Answer

Not automatically — this is a deterministic per-request cost engine, so you model those by adjusting the prices you enter. Several providers offer a batch API at roughly half price for non-real-time workloads (hours of turnaround instead of seconds); to model it, simply enter the discounted batch input and output rates. Fine-tuned and specialized models often carry different per-token rates and sometimes a separate training cost; enter the inference rates for your fine-tuned model and account for any one-time training cost separately. Because every price is a user input, the calculator adapts to batch, fine-tuned, self-hosted, or any future pricing structure without a code change.

LLM Cost Calculator

What this means

Worked example

Frequently asked questions

Component	Tokens	$/M	Cost
Input tokens	1,000	$3.00	$0.003
Output tokens	500	$15.00	$0.0075
Per request			$0.0105

Show the math