Perplexity: mixed-model agent teams beat four copies of any single model, at lower cost
perplexity_ai · x · 2026-11-18
Perplexity shared benchmark results for Multiagent.
- On Terminal-Bench 4.1, five frontier models were run solo, as 2 and 4 copies, and leading a mixed team of 4
- Every mixed team beat both its lead model solo and 4 copies of itself: +0.008 to +0.125 on Terminal-Bench 4.1, +0.015 to +0.064 on WANDR
- The Gemini 4 Carbon-led mixed team scored highest at 0.727
- Top mixed teams also cost less than 4 copies: $10.31 vs $13, and $54.50 vs $66.41 per run
Related event: Perplexity Finds Mixed-Model Agent Teams Beat Identical Model Copies(4 posts)→
More from coding & agent
- Fable 5.2 turns a fake mobile game ad into a real playable game: 70 min, 2,982 lines, $29.52 — Dan_Jeffries1 · 2026-11-18
- Google-Backed Kaggle Agent Contest Offers $400K to Crack Engineering Benchmark PLE — amankhan · 2026-11-18
- Dev snark: before running 10 agents, ask if you could even build software with one terminal — KevinNaughtonJr · 2026-11-18
- Fan-built zero-socket MCP bridge gives Project Zomboid an AI radio Game Master — Triminum · 2026-11-18
- Suno confirms its MCP includes visuals generation — suno · 2026-11-18
- Oxlint adopts tsrs backend: type-aware linting up to 7x faster, 75% less memory — cnakazawa · 2026-11-18