Perplexity CEO: diverse multi-model agent teams beat homogeneous copies on WANDR and Terminal Bench
AravSrinivas · x · 2026-11-18
Perplexity CEO Arav Srinivas argues multi-agent isn't just about cost: parallelism buys real wall-clock time savings. More notably, Perplexity's benchmarks show model diversity helps — mixed-model agent teams outperform homogeneous copies of the same agent on both performance and cost, across Terminal Bench and WANDR.
The accompanying WANDR chart (score vs median minutes per run, 1 → 4 copies) shows:
- Claude Fable 5.2: 64.3 → 48.2 min, score 0.855
- GPT-6 Astra: 37.8 → 19.0 min, score 0.80
- GPT-6.1 Sol: 33.3 → 17.3 min, score 0.77
- Claude Opus 5.5: 14.7 → 15.9 min, score 0.67 → 0.74 (slightly slower here)
He frames this as the differentiation of a multi-model platform vs single-vendor offerings.
Related event: Perplexity Benchmarks: Mixed-Model Agent Teams Win and Cost Less(3 posts)→
More from coding & agent
- Cut 25-33% of agent input token costs by giving your bot a second brain in Notion — bfrench · 2026-11-18
- "Everyone should learn to use coding agents" resurfaces as the day's most bookmarked agent advice — heyneighbor · 2026-11-18
- Databricks GA's Genie Migrate: AI converts warehouse SQL with 90%+ file accuracy — matei_zaharia · 2026-11-18
- Man builds anti-poop-eating thermal camera system for his puppy with Claude's help — nikitabier · 2026-11-18
- Perplexity: mixed-model agent teams beat four copies of any single model, at lower cost — perplexity_ai · 2026-11-18
- Perplexity launches Multiagent: a lead agent orchestrating mixed frontier-model teams — perplexity_ai · 2026-11-18