llm-prices-cn

llm-prices-cn is a price dataset repository for large language model API pricing, positioned as a daily-verified, continuously updated source of pricing for mainstream domestic and international LLMs, and comes with an MCP service that can be used for price queries and cost estimation.

1. Purpose The dataset is mainly used to centrally obtain billing information for mainstream domestic and international large language model APIs, for reference by developers, article authors, websites, or projects when they need to compare model prices or estimate invocation costs. The accompanying MCP service allows AI agents to directly query real-time prices and estimate token costs.

2. Coverage and update frequency It covers 44 models from major vendors, including OpenAI, Anthropic, Google, DeepSeek, Alibaba Tongyi, ByteDance Doubao, Baidu Wenxin, Tencent Hunyuan, MiniMax, and others. The official description states that it is verified daily to ensure that prices are authentic and reliable.

3. File structure

• prices.json: the core dataset, containing structured model details, metadata, and prices per million tokens (converted to both USD and CNY).
• sync_prices.py: a utility script used to synchronize from source application configurations and rebuild prices.json.
• CONTRIBUTING.md: contribution guidelines explaining how to report price discrepancies or request tracking for new models.

1. Data fields (prices.json)

• last_updated: last update date (YYYY-MM-DD).
• usd_to_cny_rate: the USD to CNY exchange rate used for conversion.
• pricing_unit: billing unit, fixed as per_million_tokens (per million tokens).
• models: an array of model objects, each model containing:
  • id: unique model identifier (for example gpt-5-5).
  • name: model display name.
  • vendor_id / vendor_name: vendor identifier and name.
  • country: country to which the vendor belongs (such as US, CN).
  • billing_currency: official billing settlement currency (USD or CNY).
  • input_price_usd_per_m / output_price_usd_per_m: input/output token price per million converted to USD.
  • input_price_cny_per_m / output_price_cny_per_m: input/output token price per million converted to CNY.
  • cached_input_price_usd_per_m / cached_input_price_cny_per_m: unit price when the prompt cache is hit (optional field).
  • context_window: context window size.
  • max_output: maximum single output limit.
  • modality: supported modalities (text, vision, etc.).
  • tags: model feature tags (such as flagship, reasoning, cost-effective).
  • knowledge_cutoff: knowledge cutoff time.
  • quality_score: benchmark evaluation quality score.

1. Usage

2. Directly read prices.json in the repository to obtain price data.

3. Run sync_prices.py to synchronize from source configurations and rebuild the latest prices.json.

4. Remote MCP service: endpoint https://www.llmabacus.com/api/mcp/mcp，提供工具 query_model_price(model) and estimate_cost(text_or_tokens, model), for AI agents to query real-time prices and estimate token costs.

5. Self-hosted local stdio MCP service: can be run locally, reads the accompanying prices.json, requires no network, and can be used offline. The tools provided are list_llm_prices(vendor?, currency?) and estimate_cost(model_id, input_tokens, output_tokens, currency?).

6. License and attribution The dataset uses CC-BY-4.0 (Creative Commons Attribution 4.0 International License). It permits copying and distribution in any medium or format, and also permits modification, transformation, or building upon the work, even for commercial purposes. When used, appropriate attribution must be given, and a reference to LLM Abacus or https://llmabacus.com clickable hyperlink, for example noting "Price data comes from LLM Abacus".

VII. Applicable scenarios and limitations Applicable scenarios include: comparing domestic and foreign model API prices, estimating token call costs, citing price data in projects, and enabling agents to automatically query prices and estimate costs via MCP. As for limitations: no free/paid explanation is officially provided, and related charging information is not yet available; whether a model supports a specific capability is subject to fields such as modality and tags in the dataset, and the materials do not list all model details one by one.