Every multi-agent system shipped in 2026 follows the same design: a lead agent plans the work, specialist agents execute it, and the lead combines the results. Perplexity’s Multiagent, released November 17, breaks from that pattern. Supporting agents don’t just report up — they share findings with each other, question each other’s assumptions, and flag disagreements before anything reaches the top. That difference is worth understanding, because the benchmarks show it isn’t just architectural novelty.
Peer-to-Peer, Not Hub-and-Spoke
In Perplexity’s Multiagent, a lead agent (Claude Fable 5.2) decomposes a task and assigns subtasks to up to 16 supporting agents running different frontier models in parallel. So far, that’s the standard supervisor pattern. What changes is what the workers do next: they share intermediate findings directly with each other, not just with the lead. They can question assumptions, flag discrepancies, and coordinate edits to shared files before anything gets synthesized.
Compare that to Perplexity’s own Ultra mode, where supporting agents report only to the lead. Or to OpenAI Agents SDK, LangGraph, and virtually every other production multi-agent framework in 2026 — all supervisor-pattern systems. Perplexity’s setup is closer to academic multi-agent debate, but shipped as a consumer product, mixing agents from competing model providers.
Who’s in the Room
The default configuration uses Claude Fable 5.2 as the lead, with Claude Opus 5.5, GPT-6.1 Sol, and Gemini 4 Carbon as supporting agents. Users can specify additional models in their query; Claude Fable 5.2 (as worker) and GPT-6 Astra join when explicitly requested or when a subtask calls for them. CEO Aravind Srinivas put it bluntly: “A killer for financial research.”
The model diversity is intentional. Cross-verification research consistently shows that correlated models — trained on similar data, with similar failure modes — tend to agree on the same wrong answer. Mixing Claude, GPT, and Gemini introduces genuine diversity in reasoning patterns, not just parallelism.
The Numbers That Matter
Perplexity ran mixed-model teams against four copies of a single lead model on two benchmarks. The results: on Terminal-Bench 4.1, every mixed team outperformed four copies of its lead by +0.008 to +0.125, and the top mixed team cost 0.31 per run versus ~3 for four copies of the lead. On WANDR (Perplexity’s wide-and-deep research benchmark), mixed teams gained +0.015 to +0.064 over the single-model baseline, with three of four finishing faster, at 4.50 versus 6.41.
Mixed-model teams were more accurate and cheaper than running more copies of the best available model. If you’re building agent pipelines, that finding generalizes: model diversity is a more efficient lever than model count.
The Orchestrator API
Perplexity launched the Orchestrator API the same day — currently in private preview (started November 9). It provides access to 45+ models from OpenAI, Anthropic, Google, xAI, DeepSeek, Z.AI, and Moonshot AI through one API key with unified billing and automatic provider failover. A perplexity/auto option selects the best model per request dynamically.
This is the developer-facing infrastructure underneath Multiagent. The strategic read: Perplexity is positioning itself as a neutral model routing layer — the same move Cloudflare made for web infrastructure, applied to AI inference.
Who This Is For
Multiagent requires a Perplexity Max subscription at $167 per month. It is credit-intensive — one 14-minute run burned 810 credits in testing — and Perplexity explicitly recommends it only for the most consequential tasks: forecasting, complex research, decisions where accuracy matters more than throughput.
Standard tasks and straightforward agent workflows don’t need this. High-stakes analysis where a wrong answer has real consequences is where the benchmarks show the premium earns out. The WANDR benchmark methodology is public — you can evaluate whether the accuracy gains justify the credit cost for your specific workload. Developer access to the Orchestrator API private preview is available via email request with an existing Perplexity API key.

