How TKX ranks & prices
TKX publishes daily price indices for the AI token and compute market — reproducible, source-traceable, and free via API.
The boards
Demand boards on objective public data, trust boards on transparent weighted scoring, and compute spot from live market quotes.
Router Trust
What the middlemen charge — coverage, prices, uptime, and the evidence behind each number.
Model Rankings
The models people actually use — ranked by real demand, provider coverage, and price momentum.
Apps & Agents
Which AI apps and agents burn the most tokens — the demand side of the market. Cumulative totals, not daily.
Developer Attention
Who tomorrow’s call champion may be — GitHub momentum for open-weight models.
Provider Trust
Which providers are reliable — not which models are strong.
Compute Spot
What one training run actually costs you — normalised across currency, billing grain and payment friction. Not the sticker price.
How the indices are built
Token Price Index
Equal-weighted median of completion price (USD per 1M tokens) across all paid text models listed on public aggregator catalogs. Published daily.
International GPU Rental Index
Per-GPU normalized median of verified on-demand rates (Vast.ai), reported by card type — H200 / H100 SXM / H100 NVL / A100 SXM4 / B200 / RTX 5090 / RTX 4090. USD per GPU-hour, daily.
China GPU Rental Indexcurated v1
Domestic GPU rental rates (CNY per GPU-hour) — a market international indices do not cover. v1 is a curated snapshot: per-card median across Chinese platforms' published on-demand rates (≥3 fresh quotes required), hand-verified, dated — not a live feed yet.
Token Traded Value (est.)
Estimated daily USD value of token traffic routed through OpenRouter: measured paid-tier tokens per model × that model's list rate, blended 3:1 input:output. Free-tier traffic counts toward volume but is never priced. An estimate of traffic value at list rates — not billed revenue, not global market size. Constituents stored daily so any historical value can be recomputed.
Price: mean, cheapest, and outliers
The big figure is the cheapest single measured offer for that model; the smaller figure under it is the arithmetic mean of all measured offers. Both are $/1M blended 3:1 input:output. Offers above 10x the model's median are excluded from the mean and flagged ⚠ on the detail page — still listed, still eligible to be the cheapest. The 24h % tracks the cheapest price, day over day — the same number shown above it.
Model board: volume, spend & sort order
The model board ranks by estimated 24h traded value first — paid-tier tokens × the model's list rate; free-tier traffic counts toward volume but is never priced, so free models follow, ordered by measured token volume (then provider breadth, lowest blended price). Volume and est. spend come from OpenRouter's daily usage dataset — one marketplace's measured slice, labeled with its data date. Any model with measured usage is listed even if only one provider serves it; its price columns then reflect that single source.