The TKX is live. Methodology v0.2. Baseline 100 = 2026-08-04.
The TKX scores task-specific quality and cost, on a fixed method, tracked over time.
Every model is scored across four task families — reasoning, code, extraction, classification — against its blended price. The output is a value-per-dollar score and a fair exchange rate between models, refreshed as models and prices change.
| Model class | Quality | Blended $/M | TKX value |
|---|---|---|---|
| Frontier reasoning | 96 | $20.00 | 48 |
| Workhorse general | 89 | $11.90 | 75 |
| Volume / fast | 76 | $2.38 | 100 |
| Open-weight blend | 72 | $1.65 | 96 |
Class examples only. The live TKX is published and tracked. Methodology is versioned with the benchmark (v0.2).
Crude couldn't trade globally until it was graded. Electricity, bandwidth, and computing power followed the same path: a vital, volatile input becomes a market the moment there's an agreed unit to measure it by. AI tokens have the scale, the cost, and the volatility — the TKX supplies the unit.
Answer the question every finance team is asking: are we paying a fair price per unit of intelligence? Govern spend against the index, compare every model's value on one scale, and negotiate with a reference in hand.
As computing power financializes on the world's major exchanges, contracts need a credible token benchmark to reference. The TKX is built to be that reference — published methodology, an equal-weighted basket (consumption-weighting later), and licensing for exchanges, banks, and trading desks.