classify-1 (ScaleDown)v1.7 · 10 s rule
Data: JevBench v1.7.3, published 2026-10-25 · method v1.7 (10 s rule), measured under the rule
Capability 52.3 · API offering, ranked on the API leaderboard
Last measured: 2026-10-24 · measurement release: v1.7.3
Proprietary hosted API (no weights release)
intelligence
21.8
calibration
82.8
speed
83.7
cost
50.8
Composite (secondary)
8.5 · Not ranked on this board
Cost and measurement conditions
$0.044 per 1,000 decisions (not published)
ScaleDown vendor list price, USD 0.05/M input tokens; output is free. Explicit vendor reference scaledown:classify-1 (vendor list price), with no base-model price. The frozen pricing module has no key for this API, so this row uses an explicit reference at that price. The scored cost includes the frozen scorer's usage fallback for four HTTP 504 rows: USD 0.04369407407407409 per 1,000 decisions on the common 1,485-item basis (0.67643x the USD 0.06459465517241379 cap); USD 0.082679 per 1,000 over all 1,500. Provider-reported usage alone gives USD 0.035613266 per 1,000 common items and USD 0.074679 over all items, excluding usage estimates for those failures. Cost kind mixed: four rows have estimated usage. Not a bill or quote.
p50 latency: 0.5 s
none (API)
Vendor-hosted api.scaledown.xyz/v1/scaledown API, model classify-1 (served model checked on every row; all 1,500 rows named classify-1), through first-party adapter scaledown_classify_v16; latency measured from Sandy (Germany) without API adjustment. Full S1200+P300 draw of v1.6-regular-20261012 through the unchanged run_api_full_v161.py on 24 Oct 2026; the full-set exposure gate logged all 1,200 sealed items to provider scaledown before the first request. Native probabilities on all three question types; string states sent unchanged.
Measured latency over 1,500 decisions: p50 0.505 s · p95 0.840 s
14 decisions over 10 s (10 of them answered correctly, counted as wrong).
A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.