← All leaderboards
Web search
What this board tests: web search: a query in, ranked result links out. Every provider below ran the identical corpus on 2026-09-14. The composite blends measured quality, latency p50, the track-specific cost metric and error rate. Method and corpora: how we measure.
| Rank | Provider | Composite | Recall@10 | Cost | Latency p50 |
|---|---|---|---|---|---|
| 1 | 7.82/10 | 0.67 | $0.0074 / useful result | 693 ms | |
| 2 | 7.58/10 | 0.60 | $0.00033 / useful result | 1.6 s | |
| 3 | 7.36/10 | 0.66 | $0.0076 / useful result | 679 ms | |
| 4 | 7.29/10 | 0.60 | $0.0017 / useful result | 1.3 s | |
| 5 | 6.97/10 | 0.61 | $0.0082 / useful result | 1.8 s | |
| 6 | 6.78/10 | 0.63 | $0.013 / useful result | 1.7 s | |
| 7 | 6.22/10 | 0.58 | $0.0069 / useful result | 4.3 s | |
| 8 | 5.97/10 | 0.45 | $0.018 / useful result | 2.4 s | |
| 9 | 5.74/10 | 0.42 | $0.06 / useful result | 1.1 s | |
| 10 | 5.72/10 | 0.56 | $0.0036 / useful result | 9.1 s | |
| 11 | 4.05/10 | 0.10 | $0.059 / useful result | 3.3 s |
Low scores are published alongside high ones — a ranking you can only trust if nothing gets hidden. No per-call markup, no commission from any provider. Cost figures are priced by the 2026-09-14 run's own table; the price the gateway bills today is on each provider page.