Our goal here is to measure how effective models are at using search. For simplicity, we combine publicly accepted benchmark datasets with our own WISER benchmark into a composite score. This is not meant to reflect the comprehensive evaluations Parallel runs internally. We use the following evaluation datasets: DeepSearchQA (DSQA): Google's benchmark of multi-step research questions. Humanity's Last Exam (HLE): Expert-written questions that need web-grounded reasoning. WISER (WISER): Parallel's own benchmark of real-world business research queries.

Each model answers the same 100 questions per benchmark, with and without Parallel Search Fast mode and Extract. Tool budgets are fixed; code execution is disabled. Once the tool budget is exhausted, we request a final answer with tools disabled. Reasoning settings and token limits may vary by model to reflect provider recommendations, when provided.

DSQA answers are extracted without access to the reference answer, then graded. HLE and WISER use correct-or-incorrect grading.

Search Intelligence Score: DSQA F1, HLE accuracy, and WISER accuracy, weighted equally on a 0–100 scale. F1 allows partial credit for incomplete answers and penalizes incorrect items. Scores use all 100 questions per benchmark; failed and pending tasks count as zero. Lift: the difference between a model's score with search and without. Search Efficiency Leaderboard: models at or above the median score (64.4), ranked by cost per 1,000 tasks, cheapest first. Costs include inference and estimated Search and Extract usage, excluding grading. Cost records are incomplete and may omit recovery attempts. Time per task: average seconds from question to final answer across available records, with equal weight for each benchmark. Pareto frontier: no other model is both cheaper and higher scoring. Weights: open models publish downloadable weights, under any licence; closed models are available only through an API.