Editor

I am an independent developer. I build web infrastructure for a living — mostly Laravel, Node and static pipelines — and I kept paying inference bills without any reliable way to see what a model actually costs per month before committing to it. Provider pricing pages give you one number, no cache comparison and no cross-provider spread, and the comparison articles I found were resold link farms that went stale in weeks. Price per 1M is the tool I wanted: one dataset, normalised to dollars per million tokens, refreshed by a pipeline, with the diff published instead of quietly overwritten.

What is on this site because I sat and read it

The pipeline does the arithmetic: 1,657 endpoints from 32 providers, collapsed into 45 canonical model pages, with the raw rows downloadable as CSV and JSON. What the pipeline cannot do is decide whether a figure is plausible. So the reviewing is manual, and it is what I actually do every week:

• Read the price diff that the refresh produces and sanity-check anything that moves more than a few percent.
• Open the provider's own pricing page for those moves, and drop a figure the provider no longer publishes.
• Keep a second source on the page where one exists, and show the disagreement instead of averaging it away.
• Write the 10 guides by hand around the numbers, then check each article against the dataset before it ships.

What I do not do

• No provider pays for a ranking, a placement or a favourable comparison, and no provider sees a page before it is live.
• No estimates dressed as published rates — if a provider does not publish a price, the cell stays empty.
• No copy-pasted tables from other people's comparisons; the dataset is collected and cross-checked here.

Corrections

Pricing data goes wrong: a page changes, a tier is retired, a registry lags. Corrections reach the inbox directly with the model slug, the field and a source, and are usually applied the same day. Every published page carries the review line and the date the dataset behind it was generated — currently 2026-10-28 12:57 UTC — so you can always tell how fresh a number is.

Elsewhere

Code and the dataset mirror are public on GitHub. The collection method, formulas and exclusions are documented on the methodology page and the publisher-level details are on the about page.