Hi everyone,
We’ve been building a structured pricing dataset for AI inference and wanted to share what we’ve learned with this community since many of you are making real deployment decisions every day.
The core problem is surprisingly basic: if you want to run the same model on three different providers, there’s no standardized way to compare what you’ll actually pay. One vendor charges per token, another per second, another per image, and the pricing pages are all formatted differently. Normalizing everything into comparable units turned out to be a serious data engineering challenge, especially across 51 vendors and 9 countries.
Our pipeline monitors vendor pricing pages, normalizes into standard units (per 1K tokens for text, per image for generation, per second for audio and video), and publishes weekly matched model price indexes using a chained methodology similar to how CPI or commodity benchmarks work. The goal is to isolate actual vendor price changes from composition effects when new models or SKUs enter the market.
Some patterns that might be useful to people here. The markup between buying from a model creator directly versus going through a third party platform averages around 7x on input tokens and 5x on output. Open source models price roughly 80% below proprietary alternatives for comparable tasks. Prompt caching delivers around 70% savings where it’s available, but only about 20% of SKUs currently support it. And prices across the market have been flat for several consecutive weeks, which suggests vendors are competing on model coverage and availability rather than cutting per token rates.
We also built an MCP server so AI agents can query live pricing data programmatically. If you’re building agents that need to select providers or optimize inference spend, it plugs into Claude Desktop, Cursor, or Windsurf. More details at ATOM MCP | AI Pricing Intelligence for Agents & Developers and the full benchmark indexes are free at a7om.com.
Curious to hear from anyone here running inference at scale. What pricing dimensions matter most to you when choosing a provider?
