For complex analysis, Perplexity puts different models to work together

Multiagent brings multiple models together in Computer and Teammates, with reported performance gains and high credit consumption.

Perplexity’s agents can now share findings and cross-check their analyses while working on the same task. Multiagent, a new effort level in Computer and Teammates, coordinates this work across different models. According to the official overview, Claude Fable 5.2 divides tasks among agents running Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon by default. The lead can assign up to sixteen agents while handling part of the work itself, then combine their contributions into a single response and shared deliverables.

Collaboration also addresses dependencies between analyses. In a financial model, for example, one agent can update revenue assumptions and notify the agent preparing the valuation. Exchanges also allow agents to question an assumption or flag a discrepancy for investigation. Suggested applications include financial forecasting, cross-checking financial statements against annual report footnotes, and reviewing data room documents in parallel. Agents can work on shared files and coordinate their edits.

Delegation was already available in Computer. At the Ultra effort level, supporting agents report to the lead; Multiagent adds direct exchanges between agents and cross-checking of their work. Another option, Model Council, has several models examine the same question independently before bringing together their answers and disagreements. These three approaches serve different needs: dividing up research, advancing interconnected parts of an assignment, or gathering independent perspectives.

The Multiagent option sits above Ultra in the effort settings. Web access is announced for Max and Enterprise Max subscribers, with mobile and desktop app support coming later. The launch announcement emphasizes its high credit consumption and recommends reserving it for the most important tasks.

To support its choice of different models, the team tested five models individually, as two or four copies, and leading mixed-model teams of four on Terminal-Bench 4.1. In Perplexity’s published results, every mixed team outperforms both its lead model alone and configurations using multiple copies of that model. The highest-scoring team, led by Gemini 4 Carbon, achieves 0.727. Compared with four copies of the lead model, reported gains range from 0.008 to 0.125 on this benchmark and from 0.015 to 0.064 on WANDR.

The cost comparisons concern the top mixed team on each benchmark: $10.31 per run versus approximately $13 for four copies of the lead model on Terminal-Bench, and $54.50 versus $66.41 on WANDR. On the latter test, every team of four improves on its model’s individual score, but three out of four finish sooner. These amounts describe the evaluated configurations rather than per-task prices offered to subscribers.

The lead remains the user’s point of contact throughout execution. When a follow-up request arrives, it passes the new instructions to the relevant agents so the team can continue its existing work. Users can see the agents and the models they are running; stopping the conversation thread stops the entire team.