I got one surprise AI bill that woke me up and I went looking for a simple answer: what did this model actually cost on the day I ran it?

Turns out that is weirdly hard. Provider pricing pages change quietly with no changelog. The Wayback Machine is patchy. And the obvious shortcut, today's price times last month's tokens, is just wrong the moment a provider shifts rates mid-period.

Whilst building a couple of small tools that calculate AI usage costs bugs popped up and traced back to one thing: I could not reliably say what a given model cost on a given date. Estimations are easy, knowing is hard.

So I built ai-price-index: an open, dated record of LLM API prices.

The idea is boring on purpose. Every price is bitemporal, it carries the date it became valid and the date I last re-checked it against the official source. So you can price historical usage point-in-time instead of extrapolating from today. Every price links its first-party source. A bot diffs the official pricing pages every day and flags drift, and I review every actual change by hand before it merges, because pricing is exactly the kind of data you do not want a scraper silently guessing on.

Where it stands:

• 10 providers, 86 models (Anthropic, OpenAI, Google, Mistral, xAI, DeepSeek, Cohere, Amazon Nova, AI21, Alibaba Qwen)
• CC-BY-4.0, plain JSON, DOI on Zenodo (10.5281/zenodo.20730241)
• Free, and staying free forever

If a price looks wrong or your provider is missing, that is the most useful thing you can tell me. Corrections and provider requests are open on GitHub.

Repo: https://github.com/RoninForge/ai-price-index
Browse it : https://roninforge.org/data/ai-price-index/
Demo Integration :

• Snap Shot: how much are all models on Jun 17th 2026?
• Model Price Visualization