How TKX ranks: the Trust score, explained
A ranking is only worth anything if you can see why it says what it says. This is the complete, reproducible logic behind every TKX board — no black box, no pay-for-rank.
The one rule behind everything
TKX ranks the AI token & compute market by measurement, not by claims. Every price, markup, latency and uptime figure is fetched from a source and tagged with how we know it. We never accept self-reported numbers, and we never rank on data we do not have — a dimension we cannot measure is shown as probe pending, not guessed.
The aggregator Trust score
Each live router/aggregator gets a published Trust score (0–100), a weighted blend of four dimensions:
| Dimension | Weight | What it measures |
|---|---|---|
| Coverage confidence | 35% | How many models it carries, on a logarithmic scale — 16 models is not near 315, so a thin catalogue cannot top the board on one narrow metric. |
| Price transparency | 30% | Median markup vs official list price (counted only with ≥8 comparable models). Below list scores higher. |
| Reliability | 25% | Measured uptime (requires a probe; currently pending for most). |
| Evidence quality | 10% | API-measured > official list > self-reported. |
An unmeasured dimension takes a neutral prior (0.5) — it neither rewards nor punishes — and we publish a separate completeness % so you can see how much of a score is actually measured versus assumed. Measuring more can never lower your score.
Why does a small router rank where it does?
Take a concrete case that surprises people: a web3 router carrying only 16 models. Its price transparency can be excellent (it may even price below list), but its coverage-confidence score is low because it serves so few models. So it lands mid-pack — above a large router that pads prices with a high markup, but below the high-coverage routers. The order is not an opinion; it falls out of the published weights, and you can see the four component scores on hover.
The model and provider boards
Two more boards, each with a simple, stated sort:
- Model rankings — sorted by provider breadth (how many providers serve it) then by lowest blended price. Blended price weights output, which dominates real cost:
blended = (3 × input + output) / 4per 1M tokens. - Provider trust board — sorted by models covered, then cheapest-wins count (how often a provider is the cheapest for a model), then offer count.
Evidence tags
Every data point on TKX carries one of four tags, and self-reported values are always down-ranked:
- official — the vendor's published list price.
- api — read from the provider's live public API.
- probe — measured by a TKX probe (real latency / throughput / uptime).
- self — self-reported; never treated as truth.
Reproducible by design
The full methodology is versioned (M-2026.1) and public at /methodology/; every collection run is auditable at /data/provenance.json with per-source timestamps and payload hashes; and every figure is available through the free data API. If you run a router or aggregator, you can get measured and listed automatically — see The Open Router Standard.