# Perplexity's Multiagent Sends One Prompt to Rival AI Models at Once

> Source: <https://alphasignal.ai/news/perplexity-s-multiagent-sends-one-prompt-to-rival-ai-models-at-once>
> Published: 2026-11-17 16:28:32+00:00

# Perplexity's Multiagent Sends One Prompt to Rival AI Models at Once

Perplexity's new Multiagent effort level in Computer uses a lead agent to coordinate a mixed team of frontier models, beating single-model runs on Terminal-Bench 4.1 and WANDR.

- Perplexity launches [Multiagent](https://x.com/perplexity_ai/status/2122728603909783741?utm_source=alphasignal) , a new effort level in Computer and Teammates that coordinates mixed frontier models.
- A lead agent splits tasks, workers share findings, lead merges into one final answer.
- On Terminal-Bench 4.1, mixed team led by Gemini 4 Carbon scored highest at 0.727.
- Mixed teams beat four copies of their lead by +0.008 to +0.125 on Terminal-Bench, +0.015 to +0.064 on WANDR.
- Top mixed teams also cost less: $10.31 vs ~$13 and $54.50 vs $66.41 per run.
- Credit-intensive; recommended for forecasting, dense problem solving, and consequential decisions only.

Perplexity is rolling out [Multiagent](https://x.com/perplexity_ai/status/2122728603909783741?utm_source=alphasignal), a new effort level within Perplexity Computer and its Teammates feature. A lead agent decomposes a request, assigns subtasks to workers powered by different frontier models, shares intermediate findings across the team, and synthesizes one response. The system can combine high-end general-purpose models from providers such as OpenAI, Anthropic, and Google behind a single prompt.

On Terminal-Bench 4.1 and WANDR, Perplexity’s mixed-model teams outscored teams running multiple copies of the same model while lowering the reported cost per run. The figures come from Perplexity, and WANDR is the company’s own benchmark.

## One prompt, several models

Perplexity’s [effort-level system](https://docs.perplexity.ai/docs/agent-api/effort?utm_source=alphasignal) packages a model, search configuration, reasoning budget, system prompt, and tool access into one setting. Multiagent expands that configuration into a lead-and-worker architecture:

1. The lead agent divides the request into smaller tasks.
2. Worker agents run those tasks in parallel on different frontier models.
3. The agents exchange evidence and intermediate results.
4. The lead agent reconciles the findings and produces the final response.

Perplexity Computer already coordinates specialized models for multistep work. Its model roster includes [Gemini 4 Carbon](https://cryptobriefing.com/perplexity-computer-gemini-4-carbon/?utm_source=alphasignal), which is designed for long-running tasks involving planning, reasoning, and coordination.

## Mixed teams beat cloned teams

Perplexity tested five frontier models on Terminal-Bench 4.1, a benchmark for agents completing terminal-based software tasks. Each model ran alone, in teams of two and four identical copies, and as the lead of a four-agent mixed team. Every mixed team outscored both its solo lead and the four-copy version of that lead. Higher benchmark scores indicate better performance.

| Benchmark | What it measures | Reported result | Reported cost and speed | 
|---|---|---|---|
| Terminal-Bench 4.1 | Agents completing software tasks through a terminal | Mixed teams improved scores by 0.008 to 0.125 compared with four copies of the lead model. The best configuration, led by Gemini 4 Carbon, scored 0.727. | The top mixed team cost $10.31 per run, compared with about $13 for four copies of its lead. | 
| WANDR | Agents collecting large sets of evidence-backed records from the web | Mixed teams improved scores by 0.015 to 0.064 compared with four copies of the lead model. | The top mixed team cost $54.50 per run, compared with $66.41. Three of the four mixed teams also finished faster than their solo lead while earning higher scores. | 

Perplexity separately reports that Gemini 4 Carbon scored 0.857 on WANDR as a solo model, a result the company describes as state of the art. That performance helps explain Carbon’s role as the lead in the strongest Terminal-Bench configuration.

## Why model diversity may help

Multiple copies of one model often share the same blind spots because they inherit similar training, reasoning patterns, and alignment choices. Models from separate labs may approach a subtask differently, retrieve different evidence, or catch errors another model repeats. The lead agent can compare those outputs before committing to an answer.

The division of labor also matches Google’s positioning for Carbon: assign deep reasoning, architecture, coordination, and auditing to the stronger lead, then distribute parallel work to faster or more specialized models. Repeated runs and task-level analysis would clarify how much of Perplexity’s measured gain comes from complementary errors, lead quality, or the orchestration layer itself.

## Where extra agents earn their credits

Perplexity describes Multiagent as credit-intensive and recommends it for work where broader model coverage can justify the added expense. Suitable workloads include:

- Forecasting that benefits from independent estimates and reconciliation
- Research or planning with many interacting constraints
- Large tasks that can be divided into parallel, verifiable subtasks
- Consequential decisions that warrant additional review

Casual searches, tab summaries, and quick coding questions generally offer little reason to pay for several concurrent agents. A single-model effort level remains more economical for short, low-complexity requests.

## Benchmarks meet production

Production adoption depends on whether Multiagent preserves its quality and cost advantages on a team’s own workload. Developers evaluating the feature will need to track:

- Task-level accuracy and citation quality
- Median and tail latency
- Average credits or dollars consumed per successful task
- Consistency across repeated runs
- Recovery from failed workers, tool calls, and conflicting findings
- Trace access for debugging and audits
- Data-handling terms for every model provider involved

Multiagent is rolling out through Computer and Teammates. Account availability, credit rates, model routing, and API support remain details to confirm in Perplexity’s current product documentation. Independent validation and workload-level variance also remain open questions.

## Orchestration becomes the product

Multiagent makes the orchestrated run Perplexity’s core unit of work: decomposition, parallel model calls, evidence sharing, and synthesis all happen behind the prompt box. Its practical value will depend on whether the reported cost and latency gains survive production workloads. If they do, model diversity could improve agent reliability without multiplying calls to the same system.
