Method · one ruler for all

How we measure

Every figure on a card is visible and links to its source. We publish a number only when it can be checked — and where something can’t be measured, we say so plainly rather than invent it.

Two layers per card

Layer 1 — factual axes. Crawlable specifics: API and price, output ceiling, native audio, weights licence, free tier, and more. Every cell carries a value, a trust flag, a source link and the date it was captured.

Layer 2 — output quality. Taken only from a named independent benchmark — for example an arena with human-preference Elo. Where none exists we print “no independent measure”, never a guess. Shown as an editorial Review, never an aggregate star rating.

Trust flags

official A first-party vendor surface — docs, pricing, model card, API reference or changelog.
secondary A reputable third party, a non-validating vendor surface (blog or marketing), or a figure the vendor states about itself.
unverified No source, or an unresolved conflict between sources. Shown as-is — and never used to sink a tool in a ranking.

Segments, not a single rank

There is no composite “X out of 10”. A best-for-task segment — best free, best for API, open-weights, native audio — is a gate (who qualifies, by boolean facts) plus a ranking key. A crown is awarded only when one independent Layer-2 measure can rank the whole eligible set. Otherwise we publish a shortlist ordered by a named factual key and state that basis outright. Never a crown by eye.

What we don’t measure

Plenty of models have no independent quality benchmark yet — newer releases and open-weights ones especially. We leave those blank rather than fill the gap with a guess. Figures a vendor states about itself are flagged secondary. Any listed vendor can request a dated correction at hello@answersengine.org.

Freshness

Each card shows a “data as of” date, and each cell its own capture date. The frontier — chat, image, video — is re-checked monthly; the long tail quarterly. When a new version ships, the entry moves with it. Anything going stale is labelled as such, not quietly served.

Verification log

Latest pass · 2026-07-04 — the method and its application re-checked; a sample of cards spot-checked: flags applied by protocol, Layer-2 sources valid, no invented numbers and no crowns awarded by eye.

This page is the contract. Where a card and the method disagree, the method wins — tell us and we’ll fix the card.