About LLM API Price Explorer · 關於本專案與方法論

Compare AI Model API Pricing Across Providers — Architecture, Normalization Methodology & Data Transparency.

01. Data Source & Transparency Disclaimer · 資料來源與免責聲明

Pricing data sourced from LiteLLM and NCHC TAIWAN AI RAP. Prices are indicative and may differ from official provider billing rates.

價格資料來源為 LiteLLM 及國網中心 TAIWAN AI RAP，僅供參考，實際費率請以各 API 供應商公告及帳單為準。

This platform synchronizes every 12 hours with two public sources: the official open-source BerriAI/litellm/model_prices_and_context_window.json dataset, and the NCHC TAIWAN AI RAP model list (國網中心). We do not use web scrapers, synthetic estimates, or unverified third-party pricing feeds. Absence of a model or provider in these sources does not imply that an API service does not exist.

02. Token Price Standardization · 價格標準化與精確計算公式

LiteLLM stores token rates as cost per single token (input_cost_per_token and output_cost_per_token). To enable intuitive human comparison without floating-point rounding errors, all comparable text token prices are converted using high-precision Decimal arithmetic into USD / 1M Tokens:

Input USD / 1M  = input_cost_per_token  × 1,000,000
Output USD / 1M = output_cost_per_token × 1,000,000

NCHC (one NTD price per token, same for input and output):
USD / 1M = token_price_ntd × 1,000,000 ÷ 32   (1 USD = 32 NTD)

Total Workload Cost = ((Input Tokens × Input USD/1M + Output Tokens × Output USD/1M) / 1,000,000) × Request Count
  • Missing vs Free Distinction: Records with null or missing token cost fields are strictly displayed as N/A and are never coerced to $0.00. Only records explicitly set to 0 in LiteLLM are labeled Free.
  • Special Billing Isolation: Prompt caching (cache_read_input_token_cost), cache creation, batch API discounts (*_batches), and long-context tiered pricing (*_above_200k_tokens) are stored in dedicated columns and inspectable in each model's raw spec drawer.

03. Canonical Model Normalization · 模型名稱標準化與禁止錯誤合併原則

Different API providers use different routing prefixes in LiteLLM (for example, openrouter/openai/gpt-oss-120b, deepinfra/openai/gpt-oss-120b, baseten/openai/gpt-oss-120b, and openai.gpt-oss-120b-1:0). Our normalization engine maps verified identical base weights to a canonical slug (gpt-oss-120b) while preserving every original LiteLLM ID and enforcing strict separation rules:

  • Parameter Size Isolation: gpt-oss-20b is never merged with gpt-oss-120b.
  • Quantization & Fine-Tune Isolation: Quantized variants (e.g., GGUF, FP4, AWQ) and specialized safety/fine-tuned models (e.g., gpt-oss-safeguard-120b) remain separate canonical models.
  • Creator vs Provider vs Marketplace vs Underlying Inference: We explicitly distinguish the Model Creator (e.g. OpenAI, DeepSeek, Meta), the API Provider recorded by LiteLLM, the Marketplace / Gateway (e.g. OpenRouter, AWS Bedrock), and the Underlying Inference Provider (marked Not Specified when a gateway routes dynamically without disclosing the hardware host).

04. Open Weight vs. Closed / Proprietary · 開放權重與閉源模型定義

We distinguish between Open Weight models (where model weights are publicly downloadable under permissive licenses like Apache 2.0 or MIT, or community licenses like Meta Llama) and Closed / Proprietary commercial APIs:

Open Weight

Weights are publicly released (e.g., GPT-OSS, DeepSeek R1/V3, Meta Llama, Qwen 2.5/3 open series, Google Gemma, Mistral 7B/Mixtral). Note: “Open Weight” is distinct from OSI “Open Source” when training data is not included or custom community licenses apply.

Closed / Proprietary

Weights are proprietary and accessible exclusively through commercial API endpoints (e.g., GPT-4o, OpenAI o1/o3, Anthropic Claude 3.5/3.7/4, Google Gemini 2.5 Pro/Flash, Amazon Nova).

Unknown

Models without a verified license citation in our local curated configuration (config/model-aliases.json) are conservatively labeled Unknown rather than guessing.

05. Price History · 歷史價格的記錄方式

  • Change-only records: a price is recorded when an offering first appears and whenever its rate changes. Between two records the price is unchanged, so trend charts carry each price forward as a step line up to the latest sync; delisted offerings stop at their last record. 價格只在首次出現與每次調價時記錄；兩次紀錄之間價格不變，趨勢圖以階梯線延續到最新同步日。
  • LiteLLM backfill: the 12 months before tracking began were reconstructed once from LiteLLM's own Git history, sampling one version per week. These records are tagged LiteLLM history; their date is the weekly version in which a change first appeared, so the actual change happened within the preceding week. LiteLLM 開始追蹤前 12 個月的價格，從其 GitHub 歷史版本（每週取樣）一次性還原，並標示「LiteLLM 歷史版本」。
  • NCHC: the NCHC API publishes no history, so NCHC prices accumulate from the first sync onward. 國網沒有歷史版本，從開始同步起累積。
  • No fabricated data: every point comes from a published source version; source data errors (for example a mistyped rate later corrected upstream) are kept as published rather than edited. 所有資料點皆來自來源實際發布的版本，不推估、不修改來源資料。