AI NewsModels & agentsAnnouncement
Perplexity Computer gets Multiagent, four models that check each other's work
Perplexity added a Multiagent effort level to Computer. One Claude Fable 5.2 lead runs up to 16 parallel agents. The agents use Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon.

Why it mattersOne model answering alone is cheap and often wrong. Four models checking each other on a deal diligence is what some teams would pay a junior analyst to do.
Perplexity added a new effort level to its Computer agent on 17 November. The company calls it Multiagent. The post is dated the same day. A user picks it from Computer's effort slider on the web, above Ultra.
The shape of the change is coordination. Perplexity writes that on Ultra, "one lead delegates parts of a task to supporting agents". Those agents "report back only to the lead". On Multiagent, agents "share findings directly with each other, and check each other's work". One agent can "use evidence another has found, question an assumption, or flag a discrepancy for more investigation".
The team the user runs
A lead agent runs on Claude Fable 5.2. That is Anthropic's planning and coordination model. The lead plans the task and divides the work. The supporting agents run on Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon. Those are the heavy models from Anthropic, OpenAI and Google. A task can run on up to 16 agents at once. The lead also does part of the work itself. A user can tell Perplexity which models to prefer for the job.
What this is for
The post names four jobs. The first is financial forecasting and scenario analysis. The second is auditing an annual report, with each statement checked separately. The third is a deal data room, with each document read in parallel. The fourth is complex tax planning, with year-end scenarios compared side by side. The common shape is a task that has parts. Getting each part right matters. A second model reading the same page is a useful check.
The model mix is the interesting part. A single-provider team already runs multi-agent orchestration. Anthropic's managed agents do. OpenAI's Responses API does. What Perplexity is selling is the mix. The lead that reads for discrepancies comes from Anthropic. The models that disagree come from three different labs. A reader who does not trust one lab to judge its own answer gets a judge from a different lab.
What a buyer should check
Multiagent is "available to all Perplexity Max subscribers on the web". Max is Perplexity's top plan. The post does not name a token or compute cost. A buyer should run one real task first. Read the token bill at Max pricing. Then decide whether it stays in budget for daily use.
The second check is the thing the post does not claim. There is no benchmark here. No accuracy figure. No measured reduction in analyst hours. A team using this on a deal diligence should run one test task first. Pick a task it already has an answer for. Compare the Multiagent output to the known answer. Count the disagreements between the three supporting models. That count is the useful one. It tells a buyer whether the cross-checking is doing what the pitch says.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

