Model Price Radar

io.github.alialtunarv0.1.0更新於 Oct 1, 2026

LLM prices on OpenRouter now and since 2023: history, drops, cheapest model for a job. No API keys.

已驗證STDIO僅桌面AI & MLDeveloper Tools

概覽

AI 產生的概覽

讓助理查詢 OpenRouter 上大型語言模型的目前價格、比較不同供應商、估算工作負載成本，並追溯 2023 年以來的每月價格歷史。

功能
它包裝了 OpenRouter 的公開模型目錄（約 450 個模型），以及自 2023 年 7 月起隨套件附帶的每月價格歷史，之後的月份則視需要從 Wayback Machine 取得。工具包括：find_models 依能力、上下文與價格上限篩選；estimate_cost 依工作負載成本排序最多 10 個模型；model_price_history 查看單一模型的每次價格變動；price_changes 列出季度降價、漲價與新免費模型；compare_providers 比較各供應商的價格、量化方式與可用率。另有兩個提示詞 cheapest_model_for 與 monthly_price_report。
適用情境
適合依成本或能力挑選大型語言模型、為每月呼叫量等工作負載做預算，或確認某個模型是否調價。也適合查詢具備特定能力的最便宜模型、比較供應商，以及了解價格歷史趨勢。
執行需求
以 PyPI 套件形式透過 uvx 在本機以 stdio 執行，需要安裝 uv。未宣告任何 API 金鑰或環境變數。需要存取 OpenRouter 的網路，以及（對隨套件快照之後的月份）存取 Wayback Machine。
安裝前請注意
它只讀取公開價格資料，不涉及憑證、付款或寫入操作。價格來自 OpenRouter，通常與各供應商標價一致，但不含批量與企業折扣。歷史每月僅一個資料點，因此變動只能標註在兩個月份之間；模型 id 改名會重新開始一段歷史。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Model Price Radar，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only，透過 STDIO。 STDIO 服務會啟動本機處理程序，因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

model-price-radar-mcp

What does each LLM cost today, and did it get cheaper? An MCP server over OpenRouter's public model catalog (~450 models) plus monthly price history since 2023, rebuilt from Wayback Machine copies of the catalog. Ask Claude (or any MCP client) for the cheapest model that can do a job, what a workload will cost, or how a model's price moved.

No API keys. One line to install.

You:    Did LLM prices really go down? Show me GPT-4o and DeepSeek Chat.Claude: [model_price_history ×2, price_changes]        • GPT-4o: $5 / $15 per 1M tokens until Oct 2024, $2.50 / $10 since Nov 2024.        • DeepSeek Chat changed price 12 times since mid-2024; input went from          $0.14 to $0.26 per 1M.        • Not everything gets cheaper: since Oct 2025 older open models went up          as cheap hosts dropped them, e.g. Qwen 2.5 Coder 32B input $0.04 → $0.66.

Summarized from real tool output, 1 Oct 2026.

[DeepSeek Chat price history by model-price-radar]

Why

You want to knowWithout itWith model-price-radar
Cheapest model with tools + vision + 200k contextScroll pricing pagesfind_models(needs=["tools","vision"], min_context=200000)
What 100k calls/month will cost on 5 modelsSpreadsheetestimate_cost ranks them with a "× cheapest" column
Did this model's price change?Nobody keeps historymodel_price_history since the model appeared
What moved in LLM pricing this quarterTwitterprice_changes lists drops, increases, newly free models
Which provider serves Llama cheapestCompare tabscompare_providers with quantization and uptime

How it works

mermaid
flowchart LR    C[Claude / MCP client] -->|tool call| S[model-price-radar-mcp]    S --> L[OpenRouter /api/v1/models: live prices, context, capabilities]    S --> E[OpenRouter /models/id/endpoints: per-provider prices]    S --> H[Bundled monthly history 2023→build date]    S --> W[Wayback Machine: months archived after the build]    L & H & W --> C

Price history ships inside the package (18 KB, one snapshot per month since July 2023), so history answers are instant; months archived after the release are fetched from the Wayback Machine on demand and cached. Prices are US$ per 1M tokens; "blended" assumes 3 input tokens per output token.

Install

Requires uv.

Claude Code

bash
claude mcp add model-price-radar -- uvx model-price-radar-mcp

Claude Desktop / Cursor (claude_desktop_config.json / .cursor/mcp.json)

json
{  "mcpServers": {    "model-price-radar": {      "command": "uvx",      "args": ["model-price-radar-mcp"]    }  }}

Tools

ToolWhat it does
find_modelsSearch the live catalog by name, capabilities (tools, vision, reasoning, structured outputs, audio, free), price ceilings, context
estimate_costCost of a workload (tokens per call × calls) on up to 10 models, cheapest first
model_price_historyEvery price change of one model since it appeared, with % changes
price_changesSince a month: biggest drops and increases, newly free models, models added/removed
compare_providersSame model across providers: price, context, max output, quantization, uptime

Model names are forgiving: "claude sonnet" resolves to the newest Sonnet, "llama 3.1 70b" to meta-llama/llama-3.1-70b-instruct.

Prompts: cheapest_model_for (a task → recommended model with monthly cost), monthly_price_report.

Try these

  • "Cheapest model with tool calling and vision for 50k support tickets a day? Show monthly cost."
  • "How did Claude and GPT prices change since 2024?"
  • "What got cheaper in LLM pricing since January?"
  • "Which provider serves Llama 4 Maverick cheapest, and at what quantization?"

Limits

  • Prices are OpenRouter's, which usually equal each provider's list price. Batch and enterprise discounts are not included.
  • History has one point per archived month, so a change is dated "between these two months".
  • Models are tracked by OpenRouter id; a renamed id starts a new history.

Part of the keyless MCP series

Open-source MCP servers that answer one market question each, with public data and no API keys.

ServerQuestion it answers
review-miner-mcpWhat do users hate about competitor apps and games? (App Store + Steam reviews)
pricing-time-machine-mcpHow did a SaaS pricing page change over the years? (Wayback Machine)
hn-hiring-trends-mcpWhich skills are tech companies hiring for, and which are rising? (HN Who is hiring)
model-price-radar-mcp (this one)What does each LLM cost, and did it get cheaper? (OpenRouter + price history)
launch-detector-mcpWhat is a company about to launch? (certificate transparency logs)

Development

bash
uv sync --extra devuv run pytest                              # offline tests with mocked OpenRouter + Waybackuv run python scripts/smoke_live.py        # live checkuv run python scripts/build_history.py     # rebuild the bundled history (several minutes)uv run --with rich python scripts/demo.py deepseek/deepseek-chat   # terminal demo (vhs docs/demo.tape records the GIF)

MIT © Ali Altunar

來源：README.md，提交 21dbd57

工具

0
工具後設資料尚未被收錄。

版本歷史

1
  1. v0.1.0最新Oct 1, 2026