Gemini Decision 1v1.7 · 10 s rule
Data: JevBench v1.8.10, published 2026-11-12 · method v1.7 (10 s rule), measured under the rule
Capability 78.5 · API offering, ranked on the API leaderboard
Last measured: 2026-11-11 · measurement release: v1.8.10
Proprietary hosted API (no weights published)
intelligence
67.4
calibration
89.6
speed
90.1
cost
52.5
Composite (secondary)
71.3 · Not ranked on this board
Cost and measurement conditions
$0.038 per 1,000 decisions (public list price)
Google list price USD 0.10 per 1M input tokens, output free (Gemini API paid tier; free tier, batch and caching not counted), times API-reported usage; launch list price counted from day 1; frozen pricing 1b3c52fb, page re-fetched 12 Nov 2026 15:06Z. USD 0.0382578 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0647295 over all 1,500 items); the 3 failed requests count as zero-token items.
p50 latency: 0.2 s
none (API)
Google's Gemini API (Decisions API v1beta, paid tier), model id gemini-decision-1, request-pinned (the API returns no model field). Measured on 11 Nov 2026: all 1,500 items of the 27 Oct 2026 draw, one item per request, with Google's native probabilities; latency measured from our server in Germany, network time included (p50 0.234 s, p95 0.421 s; no request over 10 s).
Measured latency over 1,500 decisions: p50 0.234 s · p95 0.421 s
0 decisions over 10 s.
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.
Model and serving disclosure
FIRST MEASUREMENT on the 27 Oct 2026 draw: all 1,500 items (1,200 sealed and 300 public), one item per request, with Google's native probabilities for the choice, yes/no and score questions; model id gemini-decision-1, request-pinned. 3 of 1,500 requests failed in transport (one HTTP 503, two network errors) and count as wrong answers; Calibration is computed on the 1,497 answered items, as for every row with failed requests. No refusals. Latency was measured from our server in Germany. SELF-REPORTED FIGURE: Google reports 92.2 % (self-reported; accuracy on the 231 public v1.2 items, vendor-run; none of those items is in this draw). Our accuracy on all 1,500 items of the 27 Oct draw: 86.3 % (a yes/no answer between 20 % and 80 % counts as unsure, i.e. wrong). ImageJevBench: requested, no date.