Because the numbers ship with the package, lookups need no key and no network, and they are exactly as current as the version you installed — the project says so and links each provider's pricing page as the source. For choosing a model for a workload that is fine; for reconciling an invoice, check the source. The comparison output prints each model's multiple against the cheapest, which is the fastest way to see how big the decision is: in the project's own example the same workload lands at 6.2x on one model and 34.6x on another.
A pricing reference that ships its own data. Rather than querying provider pages, the prices are compiled into the package at each release, so a cost comparison is a local lookup and works offline. It is a CLI and Python library first, with an MCP server included.
- Pricing for one named model — `get_model_pricing`
- The exact cost of a call given input and output token counts — `calculate_api_cost`
- The same workload priced across several models side by side, with the multiple against the cheapest — `compare_models`
- The N cheapest models for your workload — `find_cheapest_models`
- Every provider with its minimum pricing — `list_providers`
- Models found by name or filtered by provider — `search_llm_models`
- Coverage of 277 models across 47 providers, including OpenAI, Anthropic, Google, Mistral, Groq, Cohere, DeepSeek, xAI, Together AI, Fireworks AI, Perplexity, Cerebras, SambaNova, Amazon Bedrock and AI21 Labs
No API key at all. Python, and the MCP extra: `pip install "git+https://github.com/benbencodes/llm-prices[mcp]"`, which puts the `llm-prices-mcp` command on your PATH for the client to launch. The README also documents a PyPI install of `llm-prices` 0.1.44 with the `mcp` extra, a name PyPI does not have. It can also be launched with uvx. No other dependencies. Prices are frozen at each release and may drift behind provider changes, so treat a release date as the data's date.