LLM API Cost & Token Calculator

Prompt caching lets a provider reuse the processed form of a prompt prefix (system prompt, tool definitions, a long document) across requests instead of re-reading it every time. Cached input tokens are billed at a steep discount: 10% of the input price on most models, 5% on Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol, 2.5% on Claude Fable 5.1, and 2% on DeepSeek. Because the discount differs per model, a model that looks expensive at 0% cache hits can become the cheapest at 80% hits. The cache-hit slider above applies the model's cached price to that share of your input tokens.

It depends on your output volume. Coding agents are output-heavy and cache-heavy: they re-send the same repository context many times and generate a lot of code. Set the output slider to match your real usage and the cache-hit ratio to 60-80% - that is typical for agentic loops - and read the ranking. As of the latest check, Claude Sonnet 5.5 and GPT-6.1 Sol sit in the same price band ($2 in / $10 out), Gemini 3.8 Flash and DeepSeek V4 Pro are cheaper, and Claude Opus 5.5 is the premium pick at $4 / $20. Price is only one axis; benchmark quality on your own tasks before choosing.

Start from requests per day. A typical chat turn is 500-3,000 input tokens (including history) and 100-800 output tokens; a RAG query with retrieved documents is 3,000-15,000 input tokens; an agent step with tools can be 10,000-50,000 input tokens. Multiply by daily requests and by 30. For English text, 1 token is roughly 4 characters or 0.75 words; Korean and code usually cost more tokens per character. Most providers expose a usage object on every API response - log input_tokens, output_tokens and cache_read tokens for a week and plug the monthly totals in here.

Every number is copied from the provider's official pricing page and the date of the last check is stored with each model (see the footer badge and the source links). The table shows the standard tier price for prompts within the base context tier. It does not apply batch discounts (usually 50%), DeepSeek off-peak pricing (50% lower), Gemini long-context surcharges above 200K tokens, OpenAI pricing above 272K tokens, xAI pricing above 200K tokens, Claude Haiku 5.5 pricing above 100K tokens, or context-cache storage fees. Treat the result as a list-price estimate, not an invoice.

Monthly cost = input tokens × (1 − cache hit ratio) × input price + input tokens × cache hit ratio × cached input price + output tokens × output price. All prices are per 1M tokens, so the sliders are in millions. If a model has no cached price, the cache slider has no effect on it. The KRW figure is simply the USD total multiplied by the exchange rate you set (default 1,400).

Yes. Pick any two models in the VS panel (or tick two rows in the table). The panel highlights the percentage difference on input, cached input, output and total monthly cost for your workload, and generates a one-line summary you can paste into a cost review.