Skip to content
The independent index for what AI capability costs

One scoring recipe. Tracked over time.

The TKX grounds quality to cost on a published method. Baseline 100 = 2026-08-04. Current readings are on the live index.

A TKX reading above 100 means a unit of AI capability is cheaper and more efficient than the 2026-08-04 baseline of 100. A reading below 100 means it is more expensive and less efficient than that baseline.

Up = cheaper = good. Higher means cheaper and more efficient than each series' own 2026-08-04 baseline. Open vs Frontier chart height is not which basket is cheaper. Use Frontier/Open for that.

Scoring · Index construction

As of 2026-10-20 · Source: Cambent TKX

SCORING

Scoring

Composite quality uses a fixed 30 / 30 / 25 / 15 workload blend across reasoning, code, extraction, and classification. Published task scores stay on a 0 to 100 scale. We turn blended price and composite quality into one value score so models can sit on the same scale. Quality stays on a 0 to 100 task scale; price is blended dollars per million tokens. When composite quality is below 80, the value score includes a fixed penalty for that gap. A model leaves the basket when composite quality is below 70 or any task score is below 60. Models above those floors stay in the comparison. Live readings use fully scored models. Details for implementers are in the Index API.

CONSTRUCTION

Index construction

When a fully scored general model is added or removed, we carry the published series forward so the headline reading does not jump. Any membership effect is labeled separately that day.

HONESTY

Grounded, partially grounded, or provisional

We publish what we can source for the exact model you buy. When we can’t yet, we mark it plainly — Grounded, Partially grounded, or Provisional — and keep grounding more. An honest gap beats a silent guess.

Every quality score is labeled Grounded, Partially grounded, or Provisional. Grounded means published benchmarks and a sourced price for that exact model. We show the source and date. Partially grounded means some scores for this exact model are from a named source; others are not published yet. We say which. Missing scores stay unlabeled. We do not invent a score from a nearby model version. Provisional for list-only means listed, not scored yet. Other provisional estimates say grounding is in progress without inventing a score. Machine-readable coverage is on the Index API.

Scoring recipe dated 2026-08. Value-score method dated 2026-09. Membership handling dated 2026-08. Baseline 100 = 2026-08-04 on the inception basket.

Open the live index · How we forecast GET /v1/methodology

Corrections

Correction (2026-10-20): the basket member shown as 'DeepSeek V4' since 2026-08-04 was DeepSeek V3.2. Its price ($0.28 in / $0.40 out per million tokens) came from a third-party price file and was listed as coming directly from DeepSeek, but DeepSeek does not sell V3.2 on its own API. Grades and price both referred to V3.2 throughout, and the price did not change, so this member did not move any published value. From 2026-10-20 the member is named DeepSeek V3.2 and priced from AtlasCloud, a model host, at $0.26 / $0.38. We made the switch without changing the index level, so TKX and TKX Open did not move. Published values are unchanged.

Correction (2026-10-20): the prices for Claude Opus 4.8 and Claude Sonnet 4.6 (since 2026-08-04) and for the list-only Claude Opus 5 and Claude Sonnet 5 rows (since they were listed on 2026-09-06) were labeled 'Anthropic direct' but were read from Amazon Bedrock price entries, and Opus 4.8 was read from the Opus 4.6 entry. The values were the same as Anthropic's own prices (Opus 4.8 and Opus 5: $5 / $25; Sonnet 4.6: $3 / $15; Sonnet 5: $2 / $10 per million tokens), so no published value changed. Claude Opus 5 and Claude Sonnet 5 are not in any TKX basket. From 2026-10-20 all four are read from Anthropic's own price entries. Published values are unchanged.