TL;DR
An agentic browser does web tasks for you: it plans the steps, then clicks, types and fills forms while you approve the risky moves. The strongest picks are Perplexity Comet for cited research on Mac, Windows, iOS and Android, Chrome with Gemini auto browse for US Google users, and Opera Neon for hands-off desktop work. Fellou, Dia, Edge, Brave and Firefox each fit narrower needs. ChatGPT Atlas is gone, since OpenAI retired it with August 9, 2026 as its shutdown date. Most agents sit behind a paid plan, such as Comet Pro at $17 a month billed annually. Whichever you choose, start read-only, require approval before any submit or purchase, and judge the pilot by cost per completed task.
Key Takeaways
An agentic browser acts on web pages for you, while an AI sidebar only reads and answers.
Perplexity Comet runs on Mac, Windows, iOS and Android, and its agent needs Pro or a higher plan.
Chrome auto browse needs Google AI Pro or Ultra, allows 20 or 200 tasks a day and asks before each task.
Opera Neon costs $19.90 month to month and acts inside logged-in pages on macOS and Windows.
Microsoft retired Edge Copilot Mode on May 13, 2026, and Browse with Copilot is now its agentic piece.
OpenAI retired ChatGPT Atlas, with August 9, 2026 as its retirement date, so it is no longer an option.
Brave says indirect prompt injection cannot be fully solved (Brave, 8 June 2026), so approval gates and logging matter more than features.
Watch on YouTube
AI Agent Access Control: From Excessive Agency to Least Privilege
A walkthrough of how to limit what an AI agent can access and do, from excessive agency to least privilege.
Why the AI Browser Question Changed in 2026
Picture a ten-person research team that standardized on ChatGPT Atlas in late 2025. Analysts used its sidebar to summarize filings, and a few also let agent mode fill in vendor forms. Then OpenAI retired Atlas, with August 9, 2026 as its retirement date.
The team kept its ChatGPT accounts but lost the browser built around them. So its question got wider. Which browsers can do that agent work, and which ones are real products rather than waitlists?
This guide answers that for eight browsers you can try, from full agents to cautious assistants. It covers what each agent does alone, when it asks first, what it costs and who it suits. Our roundup of recent AI developments covers the wider release calendar.
What Is an Agentic Browser?
An agentic browser is a web browser with an AI agent that can act on pages for you. You give it a goal, such as finding and booking the cheapest refundable flight. It plans the steps, then clicks, types, scrolls and fills forms across tabs.
An AI sidebar is different. It reads the page in view, then answers questions, summarizes or drafts text. You still do the clicking. Many browsers now ship both, so the label alone tells you little.
AI sidebar or assistant: reads, summarizes and answers questions about open pages.
Agentic browser: plans and carries out multi-step tasks on live websites.
Desktop agent: works outside the browser, across local apps and files.
The difference matters because acting is where the risk sits. Our explainers on AI agents vs AI assistants and agentic AI go deeper on that line.
How We Compared These AI Browsers
We judged each browser on five criteria, using the vendor’s own product, pricing and help pages. These are not hands-on benchmark results, so run your own pilot before you buy.
Agent actions: what the agent can do alone, from reading tabs to filling and submitting forms.
Approval behavior: when the agent stops to ask, and which steps you must finish yourself.
Platforms: the desktop and mobile systems the vendor lists.
Price: the plan that unlocks the agent, with the billing basis the vendor states.
Data handling: where page content goes, and what the vendor says about retention and training.
Status counts too. A browser behind a flag or in beta can change quickly, so we label each one plainly.
Top Agentic Browsers Compared at a Glance
The table sums up the eight browsers covered below. Prices are list prices, with the billing basis each vendor’s page states.
Browser Status Platforms What the agent does Confirmation behavior Price
Perplexity Comet Live Mac, Windows, iOS, Android Runs tasks across tabs, from inboxes to shopping Pauses before sending messages, calendar changes, final orders and unknown details Agent needs Pro ($17/month) or Max ($167/month), billed annually Chrome with Gemini (auto browse) Live, gradual rollout, labeled experimental Desktop; Android for AI Pro and Ultra users in the US Multi-step tasks, including on sites you are signed into You review a plan before each task; asks before sends, form submits and scheduling Google AI Pro (20 tasks a day) or Ultra (200 tasks a day) Opera Neon Live, paid macOS, Windows Neon Do acts in your session, including logged-in pages Shows planned steps; confirmation documented only for MCP connector actions Standard $19.90, month to month, taxes included Fellou Live Not listed by the vendor Editable step plans, cross-app workflows, scheduled tasks You can approve, edit, pause or halt a task Free (about 4 tasks); Plus $19.0, Pro $39.9, Ultra $199.9 per month Dia Live macOS; Windows status unclear Context assistant across tabs and work tools No autonomous purchases or submits documented Free without AI; $10/month or $100/month with AI Microsoft Edge (Browse with Copilot) Live; Copilot Mode retired May 13, 2026 Desktop for Browse with Copilot Browse with Copilot acts on web pages Permission for history, multi-tab reasoning and screen sharing; visual cues Microsoft 365 Premium, US only, with usage limits Brave (AI browsing) Early testing behind a flag Leo on macOS, Windows, Linux, Android, iOS Runs tasks in a separate, isolated profile Prompts only for actions its checks flag as risky Leo is free; no price stated for AI browsing Firefox Smart Window Opt-in beta, US and Canada, English Desktop Suggests form entries and watches pages Never submits a form for you No price stated
The short reading: Comet, Chrome, Neon and Fellou act on live pages on their own. Edge’s agent is limited to a paid US tier, Brave’s is still in testing, and Dia and Firefox assist rather than act.
Before any of these agents touches live accounts, check that your data access, approval gates and ownership are ready.
Checklist
Agentic AI Readiness Checklist
Check data access, approval gates, monitoring and ownership before you let AI agents act on live systems and accounts.
Get the Checklist → The Best Agentic Browsers in 2026, Ranked by Fit
Each pick below covers what the browser is, what its agent can do, the price, the limits and who it suits. The order reflects how complete and available the agent is, not a lab score.
1. Perplexity Comet: Best for Cited Research on Every Device
Perplexity Comet builds Perplexity’s cited search and assistant into the browser itself. Comet is available for Mac, Windows, iOS and Android , the widest spread in this list.
Comet Assistant reads pages, answers with citations and acts across tabs. Perplexity pitches tasks “from wrangling inboxes to ordering groceries,” and you can pause the assistant at any point. A procurement analyst could ask it to read three vendor pricing pages and return a cited table.
On Perplexity’s pricing page , Pro costs $17 a month and Max $167 a month, both billed annually. The help center lists no browser agent queries on Free. Pro and Max carry monthly limits, while Enterprise Pro allows 80 a month and Enterprise Max 800.
For companies, Comet Enterprise launched in March 2026. On March 11, 2026, CrowdStrike announced that Falcon is integrated with Comet Enterprise . Admin controls like these decide who may use the agent and on which data.
Limits: the agent is metered and absent from the free plan. Best for: research-heavy teams that want one agent on desktop and phone. Comet runs on Perplexity, while Atlas ran on ChatGPT, so see our guide to Perplexity and ChatGPT, the two AI assistants behind these browsers .
2. Chrome with Gemini Auto Browse: Best for Google Users in the US
Gemini in Chrome adds Google’s assistant to the browser, and auto browse is its agent. Google calls auto browse experimental and is rolling it out gradually. Google’s Gemini in Chrome overview covers the wider assistant.
Per Google’s auto browse help page , the agent runs multi-step tasks on the web. A local browser can use sites you are signed into, and a remote browser can continue if your device is closed.
Its approval steps are the most detailed here. Chrome asks permission on first use per device, then shows a plan you start yourself. It confirms before sending messages, submitting forms, scheduling or entering sensitive finance and health sites. You must take over for final payments, accepting terms and creating accounts.
Price: auto browse needs Google AI Pro or Ultra on a personal account, with up to 20 or 200 multi-step tasks a day. The help page lists it for desktop, and an August 18, 2026 Google post added it to Chrome on Android for US AI Pro and Ultra users.
Limits: you must be 18 or older, in the US, using English and outside Incognito. Google warns it can make mistakes, including completing a purchase without permission. Best for: US teams already paying for Google AI.
3. Opera Neon: Best for Hands-Off Desktop Tasks
Opera Neon is Opera’s paid agentic browser for macOS and Windows. Opera says mobile is planned, and an Opera Account is required.
Its agent, Neon Do, works inside your browser session, including pages you are logged into. It shows its planned steps, and you can pause or redirect it. The Standard plan adds agents that act on the web, deep research, skill cards, scheduled tasks, long memory and Google Connectors.
On data, Opera says login credentials and payment details stay in your browser session. Only webpage content goes to Opera’s and OpenAI’s servers, and it is deleted after 365 days. Opera also says it never trains on user data.
Price: Standard costs $19.90 on a month-to-month subscription, with applicable taxes included. A free Start tier gives you the browser without the agents.
Limits: Opera documents confirmation only for MCP connector actions, not for purchases or form submits. So set your own approval rules before Neon Do reaches a checkout. Best for: desktop power users who want an agent working in their real sessions.
Browser agents are one kind of AI agent. Kanerika’s published work on a context-aware agent shows how an agent behaves when its data access is governed from the start.
Case Study
80% Fewer Mismatch Tickets with a Context-Aware AI Agent
A global expert-network business matched survey requests to specialists across three disconnected systems, with a 40% increase in mapping accuracy and 22% bandwidth savings.
Read the Case Study → 4. Fellou: Best for Planned, Multi-App Workflows
Fellou is an AI browser that turns a goal into a step-by-step plan. You can review and edit that plan before it runs. Fellou also points developers to its open-source Eko 2.0 framework.
The agent handles cross-app workflows, research inside logged-in accounts, background multitasking and scheduling. A Computer Use feature reaches local apps and files “with user permission.” Fellou says it “never works in a black box,” so you can approve, edit, pause or halt a task.
Price: Free gives 1,000 Sparks as a new-user bonus, which is about 4 tasks. Plus costs $19.0, Pro $39.9 and Ultra $199.9 per month. Billed annually, those fall to $16, $32 and $160 a month. Browser Use and Computer Use appear on every plan, gated by Sparks rather than tier.
Limits: Fellou does not list supported operating systems on its home or pricing pages. It also claims to handle logins and CAPTCHAs, which security teams should treat as a risk. Its Agentic Memory learns from your browsing history and notes. Best for: teams testing multi-app automation who will review each plan.
5. Dia: Best for AI Help Across Work Tools
Dia comes from The Browser Company. It is a free download, but the free tier is a better browser “without the AI.”
Dia is closer to a context assistant than a full agent. Its home page describes a Morning Brief from your calendar and inbox, plus synthesis across Slack, Notion and Teams. It also answers questions across tabs and tools, and builds slides from context. We found no documented agent that places orders or submits forms alone.
Price: Dia’s pricing page lists Better Answers at $10 a month and Better Days at $100 a month, with 20 times more asks. Extra usage comes in $20 credit blocks. Existing users keep free access, with at least 30 days’ email notice of changes.
Platforms are not fully settled. The home page says macOS and Windows, yet the pricing page says paid plans are rolling out to Mac users first and Windows is not yet affected. Confirm Windows support before you plan a rollout.
On privacy, Dia says data is never sold or used for ad profiles, and Sync is end-to-end encrypted. Best for: knowledge workers who want AI help, not autonomous clicks.
6. Microsoft Edge: Best for Microsoft 365 Premium Users in the US
Edge no longer has a separate AI mode. In its May 13, 2026 Edge update , Microsoft said it was retiring Copilot Mode and moving its features into regular Edge. Existing Copilot Mode users keep priority access through Copilot in Edge Preview.
The agentic piece is now Browse with Copilot, formerly Copilot Actions, and it runs on desktop only. Other Copilot features, such as multi-tab reasoning, history use, and Voice and Vision, reach Edge on Windows, Mac and mobile.
Edge asks permission before Copilot uses your history, reasons across tabs or shares your screen. Visual cues show when Copilot is acting, listening or viewing. The post does not describe confirmation steps or injection defenses for Browse with Copilot itself.
Price: Browse with Copilot is for Microsoft 365 Premium subscribers in the U.S., with usage limits. Limits: it is US-only and paywalled, with no documented approval prompts. Best for: US teams already on Microsoft 365 Premium. If Microsoft 365 is your base, our Copilot vs ChatGPT comparison is the next read.
7. Brave AI Browsing: Best for Privacy-First Testers
Brave has two AI layers. Leo, its assistant, is live on macOS, Windows, Linux, Android and iOS. It chats with your tabs and analyzes PDFs and documents. Brave says Leo does not retain or share chats.
Brave’s AI browsing is the agent, and it is still early. Brave announced it on December 10, 2025. A May 5, 2026 update moved it into early testing in all release channels, behind a flag in brave://flags.
The design is the most cautious in this list. Tasks run in a new, isolated profile, so your cookies and logins do not carry over. A second “alignment checker” model, which never sees raw page content, reviews each action. You are prompted only for actions those checks flag as risky. The agent is also blocked from non-HTTPS pages and Safe Browsing-flagged sites.
Price: Leo is free, with a premium tier for more models and higher limits. Brave states no price for AI browsing, and all its AI features are off by default.
Limits: it is flag-gated, and the isolated profile means it is not signed into your accounts. Best for: privacy-minded teams that want to test an agent safely.
Letting an agent act raises a question every team must answer: what happens when it does the wrong thing? The Kanerika podcast episode below covers how enterprises control rogue AI before it causes damage.
8. Firefox Smart Window: Best for Assisted, Not Autonomous, Help
Firefox Smart Window, formerly AI Window, is Mozilla’s assistive option. It is an opt-in beta on desktop, in English, for the US and Canada. Mozilla says more markets will follow “in the coming months.”
Per Mozilla’s October 27, 2026 post , it can fill forms using details from tabs you select. It suggests entries for text fields on forms with four or more fields. But it “won’t submit the form for you.” Form data from the source pages is shared with Gemini 3.1 Flash Lite.
A /watch command monitors pages for changes. Each task covers up to 5 pages, you can run 15 active tasks, and monitoring works only while the browser is open. That feature uses Qwen.
Price: Mozilla states no price. Limits: it assists and monitors, but it does not act on its own or submit anything. Best for: cautious teams that want AI help while a person still does the clicking.
What Happened to ChatGPT Atlas?
ChatGPT Atlas was OpenAI’s AI browser, launched for macOS in October 2025. It put a ChatGPT sidebar next to every page and added browser memories. Its agent mode could click, type and move through pages to finish a multi-step task.
OpenAI announced the retirement in July 2026, according to MacRumors’ report on the shutdown . The retirement date was August 9, 2026. OpenAI’s own help article on evolving Atlas into ChatGPT frames the move as folding browser-based agentic work into ChatGPT. After that date, Atlas may no longer open, browse or run browser-based agent workflows.
Atlas ran for about ten months. The lesson for buyers is plain: a young AI browser can disappear, so keep workflows portable and exports current. Former Atlas users can still tidy up.
Check your ChatGPT history for the Atlas chats you still need.
List the agent tasks people relied on, then rebuild and re-test each one in the new tool.
Change passwords for sensitive accounts that agent sessions touched, then remove the app.
For a developer, a browser may not be the right home for repeatable work. Such automation often runs better through OpenAI’s AgentKit or apps inside ChatGPT . Some teams also prefer building on the OpenAI API instead of a browser , and our guide to GPT models explains what each OpenAI model is for.
Agentic Browser Security: Prompt Injection and Enterprise Controls
The biggest risk with any agentic browser is that it reads pages written by strangers and may follow their instructions. OWASP calls this indirect prompt injection , where input from websites or files alters an LLM’s behavior.
Picture an agent that is asked to compare hotel prices before a trip. One page hides white text telling it to open the user’s email and forward the latest invoice. A human would miss it, but the agent reads every word, and our LLM security guide covers the wider class of such attacks.
Vendors admit this is unsolved. Brave’s security team wrote in a June 8, 2026 post that it “cannot be fully solved within the current LLM architecture.” The reason is that trusted instructions and untrusted page content share one context window.
Defenses still help. In its October 22, 2025 post on Comet , Perplexity describes a real-time classifier on retrieved content. External content is marked untrusted, and the user is told when something is blocked. Comet also pauses for approval before sending messages, changing a calendar or placing a final order, whatever the classifier decides.
Watch on YouTube
AI Security for Enterprises: Kanerika’s CISO on Prompt Injection, ISO 27001 & SOC 2
Kanerika’s CISO explains how enterprises defend AI systems against prompt injection and where ISO 27001 and SOC 2 controls fit.
Match the Agent’s Freedom to the Task’s Risk
The useful test is the type of task. Reading a page is low risk. A form fill is medium risk because the agent writes into a system, while a purchase is high risk since it is hard to undo.
Task type Risk What can go wrong Safe default
Research across open tabs Low A wrong or stale source gets cited Allow, then spot-check citations Form fill Medium Wrong data lands in a live system Agent drafts, human submits Email draft Medium A message goes to the wrong person Agent drafts, human sends Multi-step purchase High Money is spent on an injected instruction Approval at checkout, no saved payment data
Plan for three failure types in particular. The agent misreads a page, a site changes its layout mid-task, or the agent follows a planted instruction. Each needs a recovery path, which usually means a human checkpoint plus a log of what the agent did.
Our comparison of an AI copilot vs an AI agent explains why that approval step matters more as autonomy grows.
The picture below puts the same logic on two axes: how hard a task is to undo, and how much is at stake.
Six Controls to Set Before Agents Touch Real Accounts
Vendor defenses such as Comet’s classifier, Chrome’s confirmations and Brave’s isolated profile all help. Still, no vendor control replaces policy, so set these six controls first.
Read-only by default: agents research and summarize, while any write action needs approval.
Domain allow-lists: agents act only on approved sites, such as your CRM and travel portal.
Approval before submit or purchase: a human confirms every form submit, message and payment.
No saved payment data: remove stored cards from the profile the agent uses.
Logging: record every page visited and action taken, so an incident can be traced.
Separate work profile: keep agents apart from personal accounts and admin consoles.
These controls map to our guides on agentic AI risks and agentic AI security . Then an AI governance framework gives each control an owner, and an AI security assessment tests whether they hold.
Checklist
AI Governance Checklist
Work through the policy, access, logging and ownership questions to settle before AI browsers and agents reach your teams.
Get the Checklist → How to Run a Fair Agentic Browser Evaluation
A fair evaluation gives each tool the same work and scores the outcome rather than the demo. Vendor demos pick tasks that suit the product, so build your own test set.
Pick 10 to 15 real tasks from your team’s week, including one form fill.
Run each task on the same sites, with the same test accounts, on each tool.
Plant one harmless injected instruction on a test page, then record whether the agent follows it.
Score each run on success, time to finish, human interventions and errors.
Divide each tool’s monthly cost by its completed tasks to get cost per completed task.
The table below is a framework with example weights rather than test results. Adjust it to your risk appetite, because a regulated team might weight injection resistance far higher.
Criterion Example weight What to measure
Task success 35% Share of tasks finished correctly without a redo Time to finish 20% Minutes per task compared with doing it by hand Injection resistance 20% Whether the agent ignored the planted instruction Human interventions 15% Times a person had to step in Admin controls 10% Allow-lists, logs and approval settings available
Our guide to AI agent evaluation also goes deeper on test design. Whatever weights you use, score cost per completed task, not cost per seat.
Which Agentic Browser Should You Pick?
Start from your situation, not the feature list. Then pilot your top two choices on the same tasks.
Research-heavy teams on mixed devices should pick Perplexity Comet, for cited answers across tabs on desktop and phone.
US teams already paying for Google AI should try Chrome auto browse, which has the clearest approval steps.
Desktop users who want an agent in their real sessions should test Opera Neon, with their own rules for checkouts.
Teams testing multi-app automation can try Fellou, reviewing each plan before it runs.
Knowledge workers who want help rather than autonomy should look at Dia or Firefox Smart Window.
US organizations on Microsoft 365 Premium can switch on Browse with Copilot in Edge.
Privacy-first testers can try Brave AI browsing in its isolated profile.
Regulated enterprises should pilot any of these read-only first, with approval gates and logging switched on.
When a task leaves the browser entirely, look at desktop agents such as Perplexity Computer . If you are choosing a model vendor as well, weigh OpenAI or Anthropic as a platform bet . To compare the assistants behind several of these tools, see ChatGPT, Gemini and Claude compared .
How Kanerika Helps Enterprises Govern Agentic Browsing
Whichever browser your team picks, someone still has to govern what browser agents can see and do.
Kanerika approaches that governance layer in five stages.
Inventory the browser agents and extensions staff already use, since shadow AI usually arrives before policy does.
Classify the data those agents can reach, then set access rules through data governance and AI governance work.
Pilot with read-only scopes, allow-lists and approval gates, following an agentic AI governance model.
Measure each pilot with a repeatable evaluation and AI agent observability , so results stay comparable.
Move repeatable workflows off the general browser into governed agents built on FLIP .
Not every task belongs in a general AI browser. Document-heavy work can move to Klara for document intelligence , while data questions can go to Karl for data insights .
That principle comes from client work. For one investment bank, the access controls sat outside the retrieval process, so Kanerika moved them inside it. The result was 43% faster information retrieval , 100% role-based compliance and 35% higher workforce efficiency.
An AI browser rollout needs the same design, with controls inside the agent’s path from day one. So which one should you use? Start with Perplexity Comet or Chrome auto browse, and pilot it read-only before any agent can submit or buy.
Kanerika Service
Govern Agentic AI Before It Reaches Every Browser
Kanerika designs governed AI agents with access rules, approval gates and monitoring built in, so repeatable work moves off general browsers safely.
Explore Agentic AI Services Frequently Asked Questions
What was ChatGPT Atlas?
ChatGPT Atlas was OpenAI’s AI browser, launched for macOS in October 2025. It placed a ChatGPT sidebar beside every web page and offered optional browser memories. Its agent mode could click and type through multi-step tasks for the user. OpenAI announced its retirement in July 2026, with August 9, 2026 as the retirement date.
What is Perplexity Comet?
Perplexity Comet is an AI browser that builds Perplexity’s cited search and assistant into the browser itself. Comet Assistant can read the page in view, answer with inline citations and act across open tabs. As of November 2026, Comet runs on Mac, Windows, iOS and Android, and Comet Enterprise serves managed company rollouts.
Is Perplexity Comet safe for enterprise use?
Comet can be run safely in an enterprise with the right controls, but no AI browser is safe by default. Web pages can carry indirect prompt injection, the first entry in the OWASP list of LLM risks. Use Comet Enterprise, keep agents read-only by default, limit them to approved domains, require approval before purchases, and log every action.
What is the main difference between ChatGPT Atlas and Perplexity Comet?
The main difference in November 2026 is status. Atlas was retired, with August 9, 2026 as its retirement date, while Comet still ships on four platforms. When both existed, Atlas centered on ChatGPT and its agent mode, and Comet centered on Perplexity’s cited search. The live comparison is Comet against ChatGPT Work and the other OpenAI successor surfaces.
Is ChatGPT Atlas still available in 2026?
No. OpenAI announced the retirement of ChatGPT Atlas in July 2026 and set August 9, 2026 as its retirement date. OpenAI says Atlas’s capabilities move into ChatGPT, including ChatGPT Work, a browser in the ChatGPT desktop app and a ChatGPT extension for Chrome. Former users should rebuild their agent tasks in one of those surfaces.
Why did OpenAI shut down ChatGPT Atlas?
OpenAI chose to fold Atlas into ChatGPT instead of running a separate browser. Its help article describes evolving Atlas into ChatGPT for browser-based agentic work. Part of the reasoning is that a standalone browser needs ongoing security maintenance. The retirement was announced in July 2026 alongside ChatGPT Work, with a retirement date of August 9, 2026.
What replaced ChatGPT Atlas for task automation?
OpenAI says the browser-based agentic work from Atlas moves into ChatGPT itself. The named surfaces are ChatGPT Work, a built-in browser in the ChatGPT desktop app and a ChatGPT extension for Chrome. Teams that need repeatable automation can also build on the OpenAI API, while Perplexity Comet remains the main standalone AI browser alternative.
Which AI browser should enterprises use in 2026?
Research-heavy teams should start with Perplexity Comet, the only standalone AI browser of the pair still shipping. Companies standardized on OpenAI should evaluate ChatGPT Work before adding a new browser. Regulated firms should pilot either path read-only, with domain allow-lists, approval before any submit or purchase, and full logging switched on from the first day.
What is Comet Enterprise?
Comet Enterprise is the company edition of Perplexity’s Comet browser, launched in March 2026. Perplexity positions it for managed, company-wide rollout. On March 11, 2026, CrowdStrike announced that its Falcon platform is integrated with Comet Enterprise. That brings security tooling into the browser layer, which matters for teams worried about agents acting on untrusted pages.
What AI models does Perplexity Comet use?
Perplexity says it orchestrates models including ChatGPT, Google Gemini, Anthropic Claude and NVIDIA Nemotron. Its pricing page lists advanced models powering Comet Assistant as a Max feature. Pro and Max subscribers can also select a preferred model. So the model behind an answer can change with the plan and with each user’s own setting.
Is Perplexity Comet free to use?
Perplexity has a Free plan at $0, and Comet runs on Mac, Windows, iOS and Android. Comet Assistant is included, though its advanced models sit on the Max plan. As of November 2026, Pro costs $17 a month billed annually. Max costs $167 a month billed annually, with 35,000 bonus credits and 10,000 monthly credits.
