Perplexity has launched Multiagent, a top effort level for Computer that puts a lead agent in charge of up to 16 agents. The agents run on rival frontier models, work in parallel and check each other’s findings before answering. It is available now to Perplexity Max subscribers on the web.
What Multiagent Is
Multiagent sits above Ultra on Computer’s effort slider, introduced in September with four settings: Light, Standard, High and Ultra. The new tier is built for the platform’s most complex work, where a task has distinct parts that benefit from parallel analysis and cross-checking.
The lead agent is Claude Fable 5.2, which plans the task and divides it among agents running Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon. Users can name the models they want, and Fable 5.2 and GPT 6 Astra join only when requested or needed.
The difference from earlier effort levels is how the agents relate. Under Ultra, supporting agents report back only to the lead. Under Multiagent, agents share findings directly, question assumptions and flag discrepancies for further investigation before the lead assembles one response.
How the Team Works
Agents communicate during the task and coordinate edits to shared files. One agent might update the revenue assumptions in a financial model, then notify another to use those figures in its valuation. The lead stays the point of contact, bridging the team’s findings into the final response.
Perplexity pitches the tier at work where being wrong is expensive. Its examples include building financial forecasts under different scenarios, reviewing annual reports to flag inconsistent figures, testing competing explanations for a research question, and diligence on deal data rooms.
The company also lists complex tax returns and year-end planning, where agents organise income and deduction records, research tax rules and compare scenarios to prepare questions for a tax professional. For follow-ups, the lead tells the affected agents what changed and the team continues its work.
Our Read
Multiagent is Perplexity’s answer to a weakness in single-model agents: one model working alone has nobody to catch its mistakes. Putting rival models on the same job, with orders to argue, is a stronger design than asking one model to check itself. The cross-checking is the product, not a garnish.
The trade is cost and speed. Sixteen agents doing parallel work will be slower and dearer than one, and Perplexity has published no benchmarks, accuracy figures or speed comparisons for the tier. Until independent tests measure the gain, buyers pay for a process they cannot yet quantify.
The naming is worth noting too. The default team runs on models from Anthropic, OpenAI and Google, three labs that compete fiercely. Perplexity is positioning itself as the neutral ground where rival models work together, and that neutrality is now its selling point.
Why It Matters
Agent platforms are converging on the same idea: the next leap is not a smarter single model but better orchestration of many. Perplexity’s move follows Teammates, its AI-coworker launch, its Computer e-commerce expansion and the open-sourcing of Decider 1.3, its judge for AI agents.
For users, the effect is a new top gear for audits, diligence and forecasts, where an unchecked error costs money. For rivals, it raises the bar on what an agent tier must prove, since a team that cross-checks is harder to dismiss than one model with a bigger context window.
Background
Computer is Perplexity’s agent platform for research and multistep work. Ultra delegates parts of a task to supporting agents, while Model Council has several models investigate the same question independently and shows where they agree and disagree. Multiagent fills the gap: delegated work with shared, checked findings.
Source: Perplexity’s announcement and Unite.AI’s report


0 Comments