How we source facts
Every number and quote on this site links to where it came from. Nothing is ranked, and nothing is guessed.
Prices
List prices per million tokens, standard tier, from each vendor’s own pricing page. Our daily check takes a price only when two independent feeds (OpenRouter’s model list and the LiteLLM price table) agree within 2%. Anything else waits for a check against the vendor’s page, and a jump of more than 5x is held until someone confirms it.
Usage
Each model’s share of tokens on OpenRouter. It shows what people actually run. It says nothing about quality.
What the labs and tools say
Quotes from each lab’s and each tool’s own docs: how a lab describes its own models, and which model a tool uses for which job. They live in data/guidance.json and on each model, with the link and the date we checked. A lab’s words only ever describe that lab’s own models.
The quote check
A quote goes live only if its exact words are on the page it cites. The same check runs again across every quote after each data refresh, and a page whose wording changed gets flagged.
Scores
Coding score is 0–100: SWE-bench Verified where the lab or a tester published it, otherwise an estimate from other sourced results, marked est. Every score names its source. Treat scores as a rough guide.
What we never do
Never invent a number — a blank reads “—”. Never rank models or name one for you. Never favour a lab.