The TKX is live. Scoring methodology v0.2. Baseline 100 = 2026-08-04.
Construction v0.3 chains the printed TKX, Open, and Frontier levels when a fully scored general is added or dropped, so a listing does not jump the number; the composition effect is shown separately that day.
This page is the method. The live print is on the board.
The TKX is a published, versioned index of what a unit of standardized intelligence costs. v0.2 scores task-specific quality and cost. Task score is a weighted blend of curved task quality and log-scaled cost.
Four families only: reasoning, code, extraction, classification.
Headline series: TKX (general), TKX-Open, TKX-Frontier.
Composite quality: 30% reasoning, 30% code, 25% extraction, 15% classification, equal-weighted general basket. Only models scored in all four families enter that basket. Specialists rank in-family only.
We set the index at 100 when we started. Above 100 means the same unit of AI costs less than it did then. Below 100 means it costs more.
Graded means benchmark-grounded quality, mapped to 0–100 by family, for the same model variant that is priced.
Untagged scores are illustrative pending coverage.
How many prices are live versus manual, and how many scores are graded, is on the live board.
Reasoning: MMLU-Pro and GPQA.
Code: LiveCodeBench and SWE-bench.
Prices: live when a hand-verified LiteLLM key matches that variant; otherwise manual.
Construction v0.3 — chains printed TKX, Open, and Frontier levels when a fully scored general is added or dropped, so a listing does not jump the number. The composition effect is shown separately that day. Scoring unchanged. Not a re-baseline.
Scoring v0.2 — live. Baseline 100 = 2026-08-04.
Methodology changes are versioned and re-baselined. Consumption-weighting is not live.
The TKX is an index, not a router. We don’t send your traffic. Cambent recommends; you decide.