AI capability, measured with evidence
QBench Index 1.4 compares models across ten source-audited evaluations grouped into seven capability families. Every result keeps its exact model effort, harness, native score and source; missing scores are never guessed.
QBench Index 1.4 leaderboard
How QBench Index 1.4 works
Each benchmark keeps its reported 0–100 accuracy or task-success percentage. Results are averaged inside seven capability families, then the observed family means are averaged equally. Missing families are left out, not imputed and not pulled toward 50. Model effort variants are separate entries; unreported effort is never guessed.
Rank: at least 3 eligible evaluations across 3+ capability families · Evidence: full (all 7 families), broad (4–6), or limited (3) · Insufficient evidence: retained in Models, without an Index score.
Pareto frontier: open “Index vs. API price + Pareto”. The dashed frontier is on by default; the API token-price proxy is not cost per task.
Models
Models with at least one eligible result. Routers and specialist systems are identified separately. The API price catalog below also includes priced models outside the composite Index.
API price catalog
Per 1M input/output tokens. Official provider API and OpenRouter route prices are labelled separately.
| Model | Provider | Input / 1M | Output / 1M | Price source | Conditions |
|---|
Comparison
Category radar uses source-reported capability-family percentages. Empty spokes mean the category has no result for that model/effort variant.
Benchmark directory
Select a benchmark to see its model scores. The status shows whether it enters the index.