WANDR
An AI agent evaluation benchmark developed internally at Perplexity. The full name is Web Agent Navigation and Decision Reasoning. It measures a model's planning, execution, and result accuracy across real-world multi-step task chains, accounting for both performance and cost dimensions.
Timeline (last 90 days)
WANDR is an evaluation benchmark for breadth-and-depth research agents, comprising 500 tasks, each requiring 170,000 source-backed records.
In the WANDR benchmark, Claude Opus 5.5 achieved a score of 0.610.
Claude Opus 5.5's per-task cost on the WANDR benchmark is $4.13.
WANDR is an internal benchmark used by Perplexity to evaluate agent-type task performance.
A WANDR score of 0.610 does not mean the model fully succeeds on 60% of tasks; the actual partial completion rate is related to the distribution of task complexity.
Cross-vendor comparisons between internal agent benchmarks are of limited significance; longitudinal iteration comparisons within the same benchmark are more reliable.
根据官方在 WANDR 基准测试上的数据，Grok 4.6 位于性能与效率的帕累托前沿
Grok 4.6 在 WANDR 上能够匹配 Fable 5 的测试结果，而成本降低了 60% 以上
All Facts (8)
WANDR is an evaluation benchmark for breadth-and-depth research agents, comprising 500 tasks, each requiring 170,000 source-backed records.
50%UnverifiedIn the WANDR benchmark, Claude Opus 5.5 achieved a score of 0.610.
50%UnverifiedClaude Opus 5.5's per-task cost on the WANDR benchmark is $4.13.
50%UnverifiedWANDR is an internal benchmark used by Perplexity to evaluate agent-type task performance.
50%UnverifiedA WANDR score of 0.610 does not mean the model fully succeeds on 60% of tasks; the actual partial completion rate is related to the distribution of task complexity.
50%UnverifiedCross-vendor comparisons between internal agent benchmarks are of limited significance; longitudinal iteration comparisons within the same benchmark are more reliable.
50%Unverified根据官方在 WANDR 基准测试上的数据，Grok 4.6 位于性能与效率的帕累托前沿
50%UnverifiedGrok 4.6 在 WANDR 上能够匹配 Fable 5 的测试结果，而成本降低了 60% 以上
50%