llm-cost: A Lightweight C++ Tool for LLM Token Counting and Cost Estimation

This article introduces llm-cost, an open-source tool based on a single-header C++ library that helps developers quickly calculate text token counts locally and estimate API call costs for OpenAI and Anthropic models. This tool can run on Windows without complex configuration, solving the privacy and dependency issues of traditional solutions. It is a practical tool for cost control in LLM development. The original project is maintained by islaagnet27, GitHub link: https://github.com/islaagnet27/llm-cost (updated on 2026-05-24).

Cost control is a key issue in LLM development. Mainstream model providers (such as OpenAI and Anthropic) charge by tokens, but token consumption often deviates from expectations. Traditional solutions have shortcomings: online tools require sending data to third parties (privacy risks), and integrating SDKs increases project dependencies. Therefore, a localized, zero-external-dependency solution for token counting and cost estimation is needed, and llm-cost came into being.

llm-cost adopts a C++ single-header file design, with all functions encapsulated in one .h file. Its advantages include:

1. Low deployment cost: Provides precompiled .exe files, Windows users can run directly without additional toolchains;
2. High transparency: Concentrated code, easy for auditing and customization;
3. Strong privacy protection: Runs locally, no data sent to external servers, suitable for sensitive scenarios.

llm-cost has built-in tokenization rules for OpenAI (BPE algorithm) and Anthropic. The workflow is as follows:

1. Input the text to be analyzed;
2. Select the target model (OpenAI/Anthropic);
3. Split the text according to model rules and count the total number of tokens;
4. Estimate the cost using the built-in price table (updated regularly), and also consider common output ratios to give a total cost estimate, helping with budget planning.

llm-cost is suitable for various development scenarios:

• Prompt optimization: Evaluate token consumption of different prompt variants, balancing effect and cost;
• RAG system planning: Estimate the total token amount of query and context combinations, determine chunk size and Top-K retrieval count;
• Batch processing budget: Sample and estimate the total cost of processing large volumes of documents to avoid overspending;
• Model migration evaluation: Compare token count differences between different models to assist decision-making.

Usage: Windows users can download the precompiled ZIP package, extract it, and double-click llm-cost.exe to run (no command line required); developers can integrate llm-cost.h into C++ projects. System requirements: Supports Windows 10+, requires only 1GB of disk space, no additional dependencies.

Current limitations: Only supports OpenAI and Anthropic models, and the price table needs manual updates. Future improvements:

• Add support for models like Google Gemini and Cohere;
• Implement online price synchronization;
• Provide a command-line version for automated scripts;
• Develop cross-platform versions for Linux/macOS.

llm-cost embodies the pragmatic concept of "solving specific problems with minimal complexity". In today's complex LLM ecosystem, such lightweight tools are particularly valuable. It not only helps control costs but also cultivates developers' sensitivity to token consumption, making it a useful member of the LLM application development toolbox.