OpenAI leads shared-benchmark wins, but no single champion rules the latest search-API tests
Asked (summary):
Recreate and enrich the benchmark timeline from the collected results page for all players, as one common timeline with click-on/off players, and work out who the champion is.
The source page lists 8 providers; 7 of them have 62 dated benchmark results spanning 2025-02-14 to 2026-10-21 across 23 different benchmarks. Valyu appears on the page with no dated score. Scores are normalized to 0–100 for plotting, but benchmarks measure different tasks — so wins are only counted where providers met on the same benchmark.
Common timeline — all 62 results, normalized 0–100
Each point is a different benchmark. The dashed connectors only show sequence per player; they are not one consistent performance metric. Partial dates (2026-07, 2026-08) are plotted at month start and labelled at their stated precision.
Hover a point for benchmark, exact raw score, normalization and source. Click chips to toggle players. Valyu is shown greyed: no dated score on the source page.
Findings
On benchmarks with at least two represented players, OpenAI has the most head-to-head wins (5: BC-ZH, Mind2Web2, ResearcherBench, ReportBench, DRB-II), ahead of Perplexity (4), Gemini (3), You.com (2), Parallel (1), Exa (1). research.site
Latest search-API leadership is fragmented: Parallel tops BrowseComp at 94%, You.com tops FinSearchComp at 87.29%, Exa tops DRB2-20 at 45%, Perplexity tops RB-30 at 73.4%. parallel.ai
You.com narrowly leads the collected DSQA results at 86.61% versus Parallel's 86% — the two figures come from different vendor-run evaluations months apart. about.you.com
Keenable's five entries (LBC-30, AA Search, Astra-26, NEEDLE, Vals Search) are on benchmarks where this timeline holds no direct rival, so they are coverage, not competitive wins. keenableai.github.io
Cross-benchmark score means are not comparable and are not ranked here
The page's own "benchmark wins" column counts maxima on single-provider benchmarks too; the head-to-head count above excludes those.
Benchmark × player heatmap
Colour = normalized 0–100 score; bold ring = best on that benchmark among represented players. Scores across rows use different tasks and metrics (KPR, F1, match %) and are not comparable vertically.
All 62 results
Player
Date
Benchmark
Raw score
0–100
Normalization
Source
Data: 62 dated benchmark results for 7 of 8 providers, collected from research.site/table?view=charts as fetched 2026-10-22; sources include arxiv.org, github.com, kaggle.com, vendor blogs (Perplexity, Google, You.com, Parallel), Hugging Face, Artificial Analysis, Vals and the NEEDLE site — some scores are vendor-authored, not independent audits. Normalization: fractions ×100, percentages kept, other 0–100-like scores taken as reported; KPR and other benchmark-specific units are not directly comparable. Valyu has no dated score and is shown as such, not as zero.