End-to-end latency
Retrieval plus answer generation, in seconds
The headline
Comparable quality. Leaner runs.
Native
Kestrel
Total model tokens
Input + output + reasoning tokens
Transparency
Take the numbers with you.
Every chart above is computed in your browser from the 48 embedded trial records below.
View all trial-level data
| Arm | Task | Trial | Latency | Tokens | Score | Result | Failure stage |
|---|
Matched setup
Both arms received the same eight prompts and ran three times per task. Kestrel fanout used DuckDuckGo, Bing, and Yahoo concurrently.
Human-readable scoring
Quality uses semantic review, not literal keyword checks. A pass requires ≥6/8 plus minimum correctness and grounding thresholds.
Read with care
This is a small live-web benchmark, not a universal claim. Results can vary as providers, sources, and models change. The partial fallback run is excluded.