{"slug": "perplexity-brings-16-agent-ai-teams-to-iphone-and-android", "title": "Perplexity Brings 16-Agent AI Teams to iPhone and Android", "summary": "Perplexity has brought its Teammates persistent agents and its Multiagent mode to the iOS and Android apps, adding voice control for Teammates, the company announced. Multiagent uses Claude Fable 5.2 as a lead that coordinates up to 16 worker agents running Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon, and Perplexity says these mixed-model teams beat four copies of the lead model on Terminal-Bench 4.1 and WANDR, often at lower cost. Both features require a Max subscription at $200/month, or $167/month billed annually, with Multiagent consuming substantial credits while Teammates are free during a promotional period; Perplexity says roughly one-third of its most demanding Computer jobs already begin on iPhone.", "body_md": "# Perplexity Brings 16-Agent AI Teams to iPhone and Android\n\nPerplexity's always-on Teammates agents and its new mixed-model Multiagent mode are now on iOS and Android, and you can direct Teammates by voice.\n\n- [Teammates and Multiagent](https://x.com/perplexity_ai/status/2123126865829515363?utm_source=alphasignal) are now in Perplexity's iOS and Android apps, with voice control for Teammates.\n- Perplexity says about a third of its most demanding Computer work already starts on iPhone.\n- [Multiagent](https://www.perplexity.ai/hub/blog/introducing-computer-multiagent?utm_source=alphasignal) : a Fable 5.2 lead runs up to 16 agents on Opus, GPT, and Gemini.\n- Mixed-model teams beat four copies of the lead model on Terminal-Bench 4.1 and WANDR, often at lower cost.\n- [Teammates](https://www.perplexity.ai/hub/products/teammates?utm_source=alphasignal) are persistent agents with their own desktop and browser that own ongoing work around the clock.\n- Requires Max ($200/month, or $167/month billed annually). Multiagent uses a lot of credits; Teammates are free during a promotional period.\n\nPerplexity has added Teammates and Multiagent to its iOS and Android apps, giving Max subscribers mobile control over persistent agents and complex multi-agent jobs. The [mobile announcement](https://x.com/perplexity_ai/status/2123126865829515363?utm_source=alphasignal) also adds voice control for Teammates. Perplexity says roughly one-third of its most demanding Computer jobs already begin on iPhone, which helps explain the expansion beyond desktop browsers.\n\nLong-running Computer jobs execute on Perplexity’s servers while the phone acts as a control surface. Users can start, inspect, redirect, and stop work without keeping a desktop session open. The announcement covers app access and does not describe a corresponding developer API.\n\n## Persistent workers meet parallel teams\n\nBoth features run on Perplexity Computer, the company’s agent platform for research and multistep tasks such as data analysis, financial modeling, and app development. Each feature controls a different aspect of that work:\n\n- **Teammates** are persistent agents assigned to ongoing responsibilities. A Teammate can manage a growth campaign, triage bug reports, track metrics, or research a market while continuing to work in the background.\n- **Multiagent** is an effort setting for a single Computer task. It sits above Ultra and coordinates several models on jobs with connected, interdependent parts.\n\nEach Teammate runs in its own sandbox with a separate desktop, browser, browsing sessions, and local cookies. It can use the installed skills and plug-ins available to the account, with the same permissions. Those inherited permissions determine which connected services and data the agent can access. Teammates notify users when they need a decision, and mobile voice control allows users to answer, revise scope, or request a status update.\n\n## One lead coordinates up to 16 agents\n\nAccording to Perplexity’s [technical announcement](https://www.perplexity.ai/hub/blog/introducing-computer-multiagent?utm_source=alphasignal), Claude Fable 5.2 plans each Multiagent task and delegates work to agents running Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon. The lead can assign up to 16 agents while completing part of the task itself.\n\nWorker agents exchange findings during execution and edit shared files together. One agent could revise revenue assumptions in a financial model, for example, then direct another agent to use those figures in its valuation.\n\nThe lead remains the user’s single point of contact, and stopping the thread stops the full agent team. Perplexity now offers three orchestration patterns for different workloads:\n\n| How Computer’s orchestration modes differ |  |  | \n|---|---|---|\n| Mode | How work is organized | Best suited to | \n|---|---|---|\n| Ultra | The lead delegates tasks, and helper agents report their results to the lead. | Exploring several approaches to an open-ended problem. | \n| Multiagent | Mixed-model agents share findings, edit common files, and check connected work. | Large projects with interdependent parts, including data-room diligence. | \n| Model Council | Several models investigate the same question independently. Computer combines the results and identifies areas of agreement and disagreement. | Judgment calls that benefit from independent analyses. | \n\n## Mixed teams lead Perplexity’s tests\n\nPerplexity compared frontier models running alone, as teams of identical copies, and as leads of mixed-model teams. Terminal-Bench 4.1 measures performance on terminal-based tasks, while WANDR is Perplexity’s own benchmark. Higher scores indicate stronger performance.\n\n- Every mixed team outscored four copies of its lead model, with gains of 0.008 to 0.125 on Terminal-Bench 4.1 and 0.015 to 0.064 on WANDR.\n- The strongest mixed team on Terminal-Bench cost $10.31 per run, compared with roughly $13 for four copies of its lead model.\n- The strongest mixed team on WANDR cost $54.50 per run, compared with $66.41 for the corresponding four-copy configuration.\n- Every four-agent team tested on WANDR outscored its standalone model, and three of the four teams finished sooner.\n- The mixed team led by Gemini 4 Carbon produced the highest reported score, at 0.727.\n\nThese results support a testable design choice for agent-harness developers: model diversity may improve coverage when agents handle connected subtasks and share state. The figures remain vendor-reported benchmark results, and WANDR is an internal Perplexity benchmark. They do not establish production reliability, latency, or cost across other workloads.\n\n## Complex jobs carry a credit cost\n\nPerplexity describes Multiagent as credit-intensive and recommends it for forecasting, financial research, dense problem-solving, and consequential decisions. The company lists several representative workloads:\n\n1. Building financial scenarios in parallel and comparing their assumptions.\n2. Reviewing annual reports for figures that conflict across sections.\n3. Connecting contracts in a deal data room to revenue forecasts.\n4. Organizing complex tax records before professional review.\n\nLower effort settings suit quick lookups and simple coding tasks with limited dependencies. Because execution occurs on Perplexity’s servers, Multiagent primarily consumes account credits instead of local computing resources. The cited announcement provides no fixed per-task credit estimate, so consumption will depend on task scope, duration, and the number of agents involved.\n\n## Max sets the price floor\n\nAccess requires Perplexity Max, which the [pricing page](https://www.perplexity.ai/pricing?utm_source=alphasignal) lists at $167 per month when billed annually or $200 with monthly billing. Max includes 10,000 Computer credits each month, and Perplexity publishes a rate of 100 credits per dollar.\n\nThe [Teammates page](https://www.perplexity.ai/hub/products/teammates?utm_source=alphasignal) says Max plans include free Teammates usage during a promotional period, without specifying an end date. After the promotion, Teammates will consume the credits included with Max.\n\n## Launching from the phone\n\n1. Open the Perplexity app on iOS or Android and sign in with a Max account.\n2. Select a Teammate to assign work, request an update, or respond by voice.\n3. For a complex one-off job, open [Computer](https://www.perplexity.ai/products/computer?utm_source=alphasignal) and select Multiagent from the effort slider before submitting the task.\n\nPerplexity says future mixed-model systems will move toward an automated researcher, though the company has provided no timetable for that goal.", "url": "https://wpnews.pro/news/perplexity-brings-16-agent-ai-teams-to-iphone-and-android", "canonical_source": "https://alphasignal.ai/news/perplexity-brings-16-agent-ai-teams-to-iphone-and-android", "published_at": "2026-11-18 18:51:05+00:00", "updated_at": "2026-11-19 15:42:38.306791+00:00", "lang": "en", "topics": ["ai-agents", "ai-products", "large-language-models", "artificial-intelligence"], "entities": ["Perplexity", "Teammates", "Multiagent", "Perplexity Computer", "Claude Fable 5.2", "Claude Opus 5.5", "GPT 6.1 Sol", "Gemini 4 Carbon"], "also_reported_by": [], "takeaway": "Perplexity's Multiagent assigns Claude Fable 5.2 as lead over up to 16 worker agents on Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon, a mixed-model setup Perplexity reports beats four copies of the lead model on Terminal-Bench 4.1 and WANDR, often at lower cost.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/perplexity-brings-16-agent-ai-teams-to-iphone-and-android", "markdown": "https://wpnews.pro/news/perplexity-brings-16-agent-ai-teams-to-iphone-and-android.md", "text": "https://wpnews.pro/news/perplexity-brings-16-agent-ai-teams-to-iphone-and-android.txt", "jsonld": "https://wpnews.pro/news/perplexity-brings-16-agent-ai-teams-to-iphone-and-android.jsonld"}}