LLM Token Cost Calculator

The total cost of an LLM API call is calculated as: Cost = (Input_Tokens / 1,000,000) * Input_Price_Per_Million + (Output_Tokens / 1,000,000) * Output_Price_Per_Million. Here, Input_Tokens is the number of tokens in the prompt (including system and user messages), Output_Tokens is the number of tokens generated in the response, Input_Price_Per_Million is the model's cost per million input tokens (e.g., $3.00 for GPT-4o mini), and Output_Price_Per_Million is the cost per million output tokens (e.g., $12.00 for GPT-4o mini). The division by 1,000,000 converts the token count to millions, matching the pricing unit. This formula is necessary because LLM providers charge differently for input and output tokens due to the higher computational cost of generation (output) versus processing (input). The result is a monetary value (e.g., in USD) that represents the exact cost of that specific API call. For a workload, sum costs over all calls or multiply by call volume.

A 'good' cost per call depends heavily on the use case. For consumer-facing apps like chatbots, a cost under $0.001 per call is often acceptable, while enterprise document processing might tolerate $0.10-$0.50 per call. The key principle is to compare cost against the value generated (e.g., revenue per conversion or time saved). Generally, input-heavy workloads (e.g., long prompts) are cheaper than output-heavy ones because output prices are typically 3-5x higher. For budgeting, model choice is critical: smaller models (e.g., GPT-4o mini) cost 10-100x less than flagship models (e.g., GPT-4o). Always account for token caching, batching discounts, or free tiers that providers may offer. There is no single 'good' number—it's about balancing accuracy, latency, and cost for your specific task.

A common mistake is forgetting that output tokens are often priced higher than input tokens, leading to underestimates of cost for generative tasks. Another error is using character counts instead of token counts—tokens are subword units (roughly 0.75 words per token for English), so assuming 1 token per word overestimates cost. Edge cases include very long prompts (e.g., 100k+ tokens) where costs become significant, or caching that reduces input cost but not output cost. Some users ignore the per-million unit and calculate directly (e.g., 1000 tokens * $0.15 = $150), which is off by a factor of 1000. Also, forgetting to account for system messages or tool call tokens in the input count leads to undercounting. Finally, not updating prices when providers change them (e.g., price drops) can lead to outdated budgets.

Token

A unit of text processed by an LLM, roughly corresponding to a word part or punctuation mark; English text averages about 1.3 tokens per word.

Input Tokens

The tokens in the prompt sent to the LLM, including system instructions, user messages, and any previous conversation history.

Output Tokens

The tokens generated by the LLM in response to the prompt, representing the completion or answer.

Per-Million Token Price

The cost charged by an LLM provider for processing one million tokens, typically split into separate input and output prices.

API Call Cost

The total monetary cost of a single request to an LLM API, calculated from the number of input and output tokens and their respective per-million prices.