Most multi-model products are really a dropdown menu. Perplexity’s new Multiagent mode is something different: models from rival labs working on the same deliverable, passing evidence back and forth, and challenging each other’s assumptions before anything reaches you.
Perplexity launched Multiagent as a new effort level for Computer, sitting above Ultra. It builds on the effort controls the company introduced in September. Under every setting, one lead agent plans the job and hands pieces to supporting agents. What changes here is that those supporting agents now talk to each other.
How the team is built
Claude Fable 5.2 acts as the lead. It plans the task and splits it among agents running Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon. The lead can assign up to 16 agents and take on part of the work itself. Fable 5.2 and GPT 6 Astra only join as supporting agents if you ask for them or if a specific subtask calls for them.
The agents work in parallel and share findings as they go. One can pick up evidence another found, question an assumption, or flag a discrepancy for follow-up. They also coordinate edits to shared files. Perplexity’s example: one agent updates revenue assumptions in a financial model, then tells another agent to use those numbers in its valuation work.
You still talk to one point of contact. The lead merges everything into a single response, and follow-up requests get routed to whichever agents are affected. You can see which agents are running and on what models. Stopping the thread stops the whole team.
Where it fits
Perplexity points to work with distinct, interlocking parts:
- Forecasting and finance, with scenarios built in parallel and assumptions compared
- Annual report review, with statements, footnotes, and management commentary read separately, then cross-checked for inconsistencies
- Deep analysis of competing explanations for a business or research question
- Deal data room diligence across financials, contracts, and operating documents
- Complex tax returns and year-end planning, prepared for a tax professional to review
Note the framing in several of these: findings are flagged for an auditor or tax professional. That’s sensible positioning. Cross-checking between models reduces some errors, but it doesn’t make the output something you sign without review.
Ultra, Multiagent, or Model Council
Perplexity now has three ways to spread work across models, and they aren’t interchangeable. Ultra uses the same lead, but supporting agents report only back to it. Good for open-ended problems where you want several approaches tested. Multiagent lets agents share findings directly and contribute to shared deliverables, which suits big assignments with connected parts. Model Council is different again: several models answer the same question independently, and Computer shows where they agree and disagree. That’s the pick for judgment calls.
Why it matters
The interesting bet is that cross-vendor disagreement is useful signal. A single model reviewing its own work tends to repeat its own blind spots. Mixing Anthropic, OpenAI, and Google models gives Perplexity a structural argument that no single lab can easily match, since each lab’s own agent products naturally favor its own models. That neutrality is Perplexity’s real edge here.
The catch is access and cost. Multiagent is available only to Perplexity Max subscribers, on the web. Perplexity hasn’t published how much compute a 16-agent run burns, and that matters for anyone planning heavy use. Nor has it shared benchmarks showing that cross-checking beats a single strong model on accuracy.
So is this noise? Not quite. The orchestration idea is sound, and the finance and diligence use cases are where verification actually pays. But until there’s evidence on accuracy and cost, treat it as a promising workflow tool for Max users, not a proven replacement for human review.



