Perplexity Teammates and the $200 Max Plan: What a Persistent Agent Does
Always-on agents differ from chatbots in memory and initiative, but independent proof that Teammates earns its price does not exist yet.
Perplexity's Teammates, launched Nov 4, 2026, are persistent AI agents offered only on the $200-a-month Max plan. This article explains how such agents differ from chatbots, what the plan includes, what the security and productivity research says, and who might plausibly benefit.
A persistent AI agent is software that keeps a goal, a memory and access to a user's tools between conversations, and that can start work without a fresh prompt. Perplexity's Teammates, announced in a blog post dated Nov 4, 2026, is the newest product marketed this way. It is available only to individual subscribers on the $200-a-month Max plan, which costs 10 times as much as the $20 Pro plan.
The price makes the central question practical: does continuity, meaning an assistant that remembers, watches and follows up, justify the premium? As of Nov 14, 2026, the answer cannot be settled with independent evidence. The descriptions of Teammates come mostly from Perplexity and from outlets restating its announcement, and no hands-on test or benchmark of the product turned up in research for this article.
What separates a persistent agent from a chatbot
The vocabulary is still informal, and most of the explanations come from vendors and commentators rather than from standards bodies. Taskade, a company that sells agent products, defines an always-on agent as one that holds a goal, memory and app connections across conversations and starts work when an event, a schedule or its own judgment triggers it. That contrasts with the familiar chatbot, which waits for a prompt, answers and, in the common case, keeps nothing afterward.
The Institute of Product Management, a training provider, describes the architecture behind such tools. In its account, a persistent agent runs on a remote computer with its own file system, tools, processes and memory, so it can keep working after the user closes the window. That analysis is expert commentary, not peer-reviewed research.
Four ingredients that define the category
Across the explainers reviewed, the same features recur. They are listed here as Jaroha's synthesis of those descriptions, not as an official definition.
- Memory that survives between sessions, so past work and feedback inform the next task.
- Standing access to tools and accounts, such as messaging, code repositories or a browser.
- Triggers other than a human prompt, such as an incoming message, a schedule or the agent's own judgment.
- Background execution on a remote machine, so work continues while the user is away.
A spectrum, not a switch
A Substack guide from AI Regeneration proposes four categories. Chatbots answer and forget. Agents such as Codex and Claude Code handle complex tasks, but only when asked. Companions remember the user. Always-on agents combine capability, memory and background running. The author's framing is opinion, but it shows why the boundary is blurry: scheduled tasks and on-request agents sit between the poles.
The industry is converging on the idea. Business Standard reports that OpenAI, Meta and Google are all moving toward persistent agents, including a Google product called Gemini Spark that it describes as a 24/7 agent. The competitive pressure helps explain why Perplexity framed the launch as it did, though the brief does not show how many customers want the capability.
What Perplexity actually launched
According to Perplexity's blog post, teammates are personal agents for long-running work. A teammate coordinates sessions of Computer, Perplexity's existing agent, along with other teammates. It plans the work, checks the results, returns finished output for the user's review, remembers earlier work and applies feedback.
Perplexity's help center adds operational detail. Teammates can watch connected tools such as Slack, GitHub and Linear, begin work when something arrives and follow up, for example by responding to review comments on a code pull request. A user can create several teammates with separate responsibilities. Each reuses Computer's existing tool and project permissions and runs in its own environment with separate browser sessions.
Engadget reports that teammates work in the background around the clock and notify the user when a decision is needed, with each one getting its own sidebar entry in Computer. According to a secondary summary of Perplexity's examples from DataStudios, the use cases include an ads manager that stays inside spending caps, an engineer that opens pull requests with rollback plans, and a legal research assistant that drafts redlines. That is a single secondary source, and the examples describe intent rather than measured results.
A research preview with limits
The help center labels Teammates a research preview and says it requires an individual Max subscription; Enterprise Pro and Enterprise Max plans do not include it by default. The phrase matters because "always-on" is largely marketing shorthand, and the company itself signals that the feature is still being shaped. Perplexity says it will expand access and features in the coming weeks, without giving dates.
Sources disagree on where it runs. Perplexity's blog says the web first, with iOS soon. Engadget describes web and iOS availability, and Digital Trends reported around Nov 13–14 that it had reached the Mac app. The most likely explanation is a fast-moving rollout, so the platform list is best treated as a snapshot as of mid-November.
CEO Aravind Srinivas framed the change as a move from writing prompts within a session to handing over high-level goals, according to AI Chat Daily's account of his statements. That is a company description of the intent, not an evaluation of how well the product delivers on it.
What the $200 buys, and how credits work
Teammates is not a standalone $200 product. The base agent, Computer, launched Feb 25, 2026, for Max subscribers only, according to Fello AI and VentureBeat. Fello's later page says Computer is now also available on Pro and Enterprise plans, though many articles still repeat the old Max-only claim. The accurate framing is that Computer is the cheaper layer and Teammates is the Max-only layer on top.
CloudZero's pricing guide, dated around September 2026, says Max includes everything in Pro plus 10,000 Computer credits a month, frontier models, Model Council and a feature called Brain, and 10 times the limits on uploads and asset generation. TechJack Solutions, citing the October 2026 pricing page, adds a 35,000-credit bonus, and FreeAcademy describes that bonus as one-time. These figures come from secondary pages that say they read Perplexity's pricing; the pricing page itself was not checked directly.
| Plan | Monthly price | Annual price | Teammates |
|---|---|---|---|
| Free | $0 | – | No |
| Pro | $20 | $200 (about $16.67 a month) | No |
| Max | $200 | $2,000 (about $166.67 a month) | Yes (research preview) |
| Enterprise Pro | From $40 per user | – | Not by default |
| Enterprise Max | $325 per user | $3,250 per user | Not by default |
Metering and Jaroha's arithmetic
Computer runs on credits at a stated rate of 100 credits per dollar, according to Perplexity's credits guide as relayed by Second Talent. By Jaroha's calculation, the 10,000 monthly credits are therefore worth roughly $100 of metered use, about half the $200 fee. Paying monthly costs $2,400 a year ($200 × 12); the annual option is $2,000 according to TechRepublic, while FourWeekMBA describes $167 a month billed annually, which would be $2,004. The gap is trivial, but it shows how loosely the secondary sources treat the numbers.
The unknown that matters most is how many credits a teammate burns. According to FourWeekMBA's reading of the launch material, Perplexity gives no usage, customer or revenue figures and does not say how many Computer credits a teammate consumes. ClaudeGPT.si reports that teammate usage is free during a promotional period and then draws on Max credits, with no separate price; that is a single source and the end date could not be confirmed.
GeoToolbox, again as a single source, says every Max account has a monthly credit spending limit that defaults to $200 and can be raised to $5,000, and that running tasks pause when the limit is reached. If accurate, heavy use of always-on agents could cost well beyond the plan fee or stall mid-task. That claim needs verification against Perplexity's own documentation.
Rivals and pricing parity
Fello's comparison puts Max at the same $200 top price as ChatGPT Pro and Claude's top tier, and says Perplexity bundles 10,000 Computer credits for an agent feature the others sell separately. BigGo claims the opposite, that rivals fold comparable capabilities into cheaper tiers, but offers no source. Neither claim was checked against the vendors' own pricing pages, so a price comparison remains unresolved. AI Chat Daily lists OpenAI workspace agents, Microsoft's Copilot Cowork, Grok Bot and Anthropic's Claude Cowork as competitors, noting that the OpenAI and Microsoft products target organizations.
Reliability and security: what research says about agents in general
An agent that reads web pages and email, holds credentials and acts between sessions presents a larger attack surface than a chat window. Aurascape, a security vendor, makes that argument: a chatbot touches data only when asked, while an always-on agent holds standing access and memory. The firm also relays a 2026 Cloud Security Alliance survey reporting that 82% of organizations have unknown AI agents and 65% have had agent-related incidents; the original survey was not located, so those figures are unverified.
Prompt injection
A 2026 review posted to arXiv, a preprint server where papers are not peer-reviewed, says language models do not reliably separate trusted instructions from untrusted text. That weakness matters more for agents because reading outside content is part of their job. The review concludes that the studies it surveyed "establish partial risk reduction, not a general solution."
Numbers vary with the test. The same review reports attacks succeeding roughly 1 time in 4 against a ReAct-prompted GPT-4 agent in a 30-agent benchmark of 1,054 cases. The investor essay from 1kx describes an indirect prompt-injection competition with 464 participants, about 271,000 attempts and 13 models, in which per-attempt success rates ranged from 0.5% to 8.5%. Those two results come from different setups and should not be compared, and the essay is a secondary relay of a study whose original was not traced.
None of this tests Teammates. It describes a class of risk, not Perplexity's particular exposure. Perplexity says teammates run in isolated environments, and VentureBeat reports that the company has argued its cloud-based approach is safer than local-access models; that is a company claim the brief could not test.
Consistency and productivity
A benchmark preprint called Claw-Eval found that injected tool failures mainly erode consistency rather than peak capability. Models usually found at least one successful path to a goal but rarely succeeded on all three tries. For a recurring background job, the gap between sometimes and always is the practical issue.
Productivity evidence is mixed. A randomized trial of 16 experienced open-source developers across 246 tasks found that early-2025 AI tools increased completion time by 19%, even though participants expected a gain. A later METR study produced estimates consistent with speedups from newer agents, but participant selection and unreliable time measurement prevented a firm estimate. Both findings reach this article through a secondary arXiv preprint, concern coding work and say nothing directly about Teammates or non-coding tasks.
Who might benefit, and who probably would not
The following is interpretation, not tested evidence. The product description suggests the best fit is recurring, well-defined, digital work whose output a person can review: competitor monitoring, ticket triage, bug triage and reporting, which are the use cases Perplexity names. Someone who would pay for Max anyway, for the higher limits and credits, gets Teammates as an addition rather than the sole reason to upgrade.
Casual users have weaker grounds. Tasks that are one-off, or high-stakes with no review step, play to neither the memory nor the background execution. Commentators from Lindy, which sells AI agents and so has a conflict of interest, and from Enzo Soler's personal site both advise a narrow pilot or not upgrading just for this feature; both are opinion.
Oversight cuts both ways
WebUtility's summary says Perplexity's materials stress review points, escalation and the ability to inspect or take over an agent's desktop. Human checkpoints can limit damage from errors or injected instructions. They also reduce the labor the agent saves, so the value of a teammate depends on how much checking its output requires, a figure nobody has published.
Paying for early access
Computer's history offers a clue about pricing trajectory. It began as a Max-only feature and has since reached cheaper plans, and one commentator cited in the research notes that Perplexity has a record of rolling features down after an initial period. If Teammates follows the same path, a $200 subscription now may partly buy early access, though Perplexity has given no date or commitment for wider availability.
What it means, and what to watch
Teammates is best understood as an experiment in how much initiative people will hand to software, priced at a point that filters for heavy users and the curious. The concept of continuity is well defined, but the value claim rests on a company's description and a few vendor-flavored guides. Until someone runs a fixed set of recurring tasks for several weeks and reports the cost, error rate and review time, the sensible reading is that $200 is a bet on early access.
Three developments will clarify the picture: Perplexity publishing credit consumption for teammates, independent testing of reliability and safety controls, and whether the feature moves to cheaper plans. The persistent-agent idea itself seems unlikely to fade, given the interest from OpenAI, Meta and Google. What remains open is whether any of these agents earns trust, at any price, with the standing access the design requires.
Frequently asked questions
What is a persistent AI agent?
It is software that keeps a goal, memory and connections to tools between conversations and can start work on a trigger, such as an incoming message or a schedule, without a new prompt. Descriptions often add that it runs on a remote computer so it can work after the user disconnects. The definitions come mainly from vendors and commentators, and the line between agents and chatbots is a spectrum.
How is Perplexity Teammates different from a chatbot?
According to Perplexity, a teammate remembers past work, applies feedback, watches connected tools such as Slack, GitHub and Linear, and returns finished work for review. A chatbot generally responds only when prompted. Teammates is built on top of Perplexity's Computer agent and is labeled a research preview.
How much does Perplexity Teammates cost?
Teammates has no separate price. It requires an individual Max subscription, listed by secondary sources at $200 a month or $2,000 a year. Max reportedly includes 10,000 monthly Computer credits, worth roughly $100 at 100 credits per dollar, and how many credits a teammate uses has not been disclosed.
Is the $200 Max plan worth it?
No independent evidence yet shows that Teammates saves enough time or money to justify the price. Recurring, well-defined digital work that can be reviewed, and a need for Max's other features, make the case plausible; casual or one-off use makes it weaker. That is interpretation, and commentators in the research suggest a narrow pilot first.
Are always-on AI agents safe?
They carry a larger risk surface because they hold standing credentials and read untrusted content. Research summarized in a 2026 arXiv preprint review says prompt injection has only partial mitigations, not a general solution. No study examined here tests Teammates, and Perplexity's claims about isolated environments could not be verified.
Sources (30)
- Teammates are agents for complex, ongoing work , Perplexity (2026-11-04)
- Get started with Perplexity teammates , Perplexity Help Center
- Perplexity's 'teammates' are the company's AI agents for daily work tasks , Engadget
- Perplexity Teammates bring autonomous AI agents to the Mac app , Digital Trends
- Perplexity Teammates Launches: How Its $200-a-Month AI Agents Work , TechRepublic
- Perplexity Launches Teammates for $167-a-Month Max Users , FourWeekMBA
- Perplexity launches 'Computer' AI agent that coordinates 19 models, priced at $200 a month , VentureBeat
- Perplexity Pricing 2026: Plans, Limits, API Costs & Comet , Fello AI
- What Is Perplexity Computer? Pricing, Credits and Review , Fello AI
- Perplexity pricing in 2026: Free vs. Pro vs. Max, and who should pay , CloudZero
- Perplexity Pricing 2026: Plans, Pro, Max, and Is It Worth It? , GeoToolbox
- Perplexity Pricing: Pro, Max & API Rates (2026) , TechJack Solutions
- Perplexity Free vs Pro vs Max (2026): Prices and Limits Compared , FreeAcademy
- Perplexity AI Features and Capabilities in 2026: The Complete Guide , Second Talent
- Perplexity Pricing: Plans, Credits, and the Real Cost , Lindy
- Perplexity Teammates: work agents on Max only , Enzo Soler
- Perplexity Teammates: always-on agents for Max users , ClaudeGPT.si
- Perplexity Teammates: always-on AI agents for Max at $200 a month, credits and costs , DataStudios
- Perplexity launches Teammates, persistent AI agents for Max subscribers , AI Chat Daily
- Perplexity Teammates Launch as AI Agents for Ongoing Work on Max , WebUtility.io
- Perplexity Debuts Always-On AI 'Teammates' for Work, Locked Behind $200-a-Month Max Plan , BigGo Finance
- Always-On AI Agents: Dots vs Muse vs Grok Bot (2026 Guide) , Taskade
- [Always-on agents] Part 1: A guide to the new wave of AI , AI Regeneration (Substack)
- Chatbots to always-on AI: Why OpenAI, Meta, Google want persistent agents , Business Standard
- Persistent AI Agents: The Product Strategy for Always-On Agent Architectures , Institute of Product Management
- Risks of Always-On AI Agents in Messaging Apps , Aurascape
- Trustworthy Agentic AI: Failure Modes, Mitigation Strategies, and a Lifecycle Framework for Autonomous LLM Systems , arXiv (preprint)
- Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents , arXiv (preprint)
- Resource Constraints and Performance in Agentic AI Systems , arXiv (preprint)
- Security and Trust: The Bottleneck of Deploying Agentic AI , 1kx