Data · Edition 2026.10
Every number, as the tools wrote it
The files behind every page of this site, unchanged. They are also in the repository under results/, with the tools that wrote them.
Files
- /api/v1/leaderboard.jsonOne row per model: score, accuracy, calibration, speed and memory on each device
- editions/2026.10.jsonThe edition: its runs, models, scores and flags
- runs/2026-10-14-latency-rtx5090-cuda.jsonSpeed and memory on the GPU
- runs/2026-10-14-latency-rtx5090-cpu.jsonSpeed and memory on the CPU
- runs/2026-10-14-public-benchmark-rtx5090-cuda.jsonPublic benchmark: per model and per subset
- runs/2026-10-14-typed-decisions-rtx5090-cuda.jsontyped-decisions: per model, type and workflow
Per-question outputs (every request and every answer, about 6,000 per model) are kept with the edition’s raw files and will be published as a dataset. The JSON API is static: it changes only when an edition does.