The naive version — average price per token this quarter against last — mostly measures which models were released, not whether serving got cheaper. Chaining matched models compares the same thing to itself over time, which is how price indices are built in every other market and is the only version that supports a claim about direction. The per-model cross-vendor comparison is the practically useful half: the same open weights are served at meaningfully different prices, and nobody checks because checking means visiting eight pricing pages.
An AI inference pricing intelligence service: model search across many vendors, cross-vendor price comparison for the same model, vendor catalogues, and a price index with market-level indicators.
- compare_prices works per model or model family across vendors, which is the comparison that matters — the same weights cost different amounts depending on who is serving them.
- get_index_benchmarks returns a chained matched-model price index, meaning the comparison is like-for-like over time rather than an average dragged around by which models exist.
- get_model_detail combines technical specifications with pricing across every vendor serving it.
- search_models filters across the tracked vendors and their SKUs, and list_vendors returns who is covered.
- get_vendor_catalog returns one vendor's full range with modalities and pricing.
- get_market_stats and get_kpis return aggregate counts and market-level performance indicators.
Nothing named.
One command — npx -y atom-mcp-server