LLM API Cost Calculator

How LLM pricing actually works

Every major provider charges per token, quoted per million tokens, with input and output priced separately. The asymmetry between the two is the single most important thing to understand: output tokens typically cost three to five times what input tokens cost, and on the largest models the gap is wider still.

That changes how you design. A feature that feeds in a long document and returns a short classification is cheap, even though the prompt is enormous. A feature that takes a one-line instruction and generates a full article is expensive, despite the tiny prompt. When people are surprised by an AI bill, it is almost always because the output side grew without anyone watching it.

Two other mechanics move the number materially. Prompt caching, offered by most providers, charges a reduced rate for repeated prefix content — if every request shares the same long system prompt, caching can cut input cost substantially. Batch processing typically offers around half price for work that tolerates a delay of several hours. Neither is modelled here, so treat the result as an unoptimised ceiling.

Estimating your token counts realistically

The calculator is only as good as the token figures you feed it, and people consistently underestimate the input side. Some things worth counting that are easy to forget:

• The system prompt is sent every time. A 2,000-token system prompt on 100,000 calls a month is 200 million input tokens before any user content.
• Conversation history compounds. In a chat feature, turn ten resends turns one through nine. Average cost per turn grows through the conversation; the mean across a session is far higher than the first turn.
• Retrieved context is usually the largest component. A RAG system injecting five documents of 1,000 tokens each is spending 5,000 input tokens per call on retrieval alone.
• Tool and function definitions count. A rich tool schema can run to several thousand tokens, resent on every request.
• Retries are real calls. If five per cent of requests fail validation and retry, add five per cent.

Use the token counter on a representative prompt rather than guessing. For the output side, generate a few real responses and measure them — models are usually more verbose than people expect, which is why explicit length instructions pay for themselves.

Reading the comparison sensibly

The cheapest model in the table is rarely the right answer. A small model that needs two attempts, or that requires a longer prompt with more examples to reach acceptable quality, can easily cost more in practice than a capable model that gets it right first time. The useful comparison is cost per successful outcome, not cost per call.

The pattern that works well in production is routing: send the straightforward majority of requests to a small fast model, and escalate only the cases that fail a confidence check to a larger one. If ninety per cent of traffic can be handled by a model costing a tenth as much, the blended rate approaches the cheap model's price while quality tracks the expensive one's.

One more caution. The prices here are reference values recorded on a date shown in the tool, and they are editable precisely because they go stale. Before you commit a budget to a spreadsheet, copy the current figures from the provider's own pricing page into these fields.

How to use the LLM API Cost Calculator

1. Enter your token counts

   Input tokens per call and output tokens per call. If you are unsure, measure a real prompt with the token counter first.

2. Enter your call volume

   Requests per month. Include retries, and remember that background jobs often generate far more calls than user-facing traffic.

3. Check the prices are current

   The date the reference prices were recorded is shown in the tool. Every price field is editable — paste in the provider's current figures.

4. Compare the totals

   Look at monthly cost, but also at cost per call. A model that is ten times cheaper per call may still be the wrong choice if it needs two attempts.

Frequently asked questions