OpenAI has quietly set an execution date for one of the three products at the center of the AI browser race. ChatGPT Atlas, the desktop browser OpenAI launched with a splashy demo in October 2025, stopped functioning as a browser on August 9, 2026, according to OpenAI’s own help documentation. Meanwhile, Perplexity’s Comet just landed on Android on August 19, 2026, branding itself the first agentic AI browser built for mobile. Google, for its part, already euthanized Project Mariner back on May 4, 2026, folding its web-navigation engine into Gemini Agent and a feature called Chrome Auto Browse. Three companies, three very different outcomes, all inside a single year.
That timeline matters because it changes the calculus for anyone deciding which AI browser to build a workflow around in late 2026. This is not a hypothetical “which is best” exercise. One of the three contenders is being actively wound down, and its users need somewhere to land. This guide compares Perplexity Comet, ChatGPT Atlas, and Google’s Gemini Agent (plus Chrome Auto Browse) on specs, pricing, benchmarks, security posture, and real-world task performance, and it ends with a migration path for anyone still running Atlas past its shutdown date.
What Is an Agentic AI Browser, and Why Every Chatbot Maker Built One
An agentic AI browser is a normal Chromium-based browser with a control layer bolted on top that can read a page’s DOM, plan a multi-step sequence of clicks and form entries, and then execute that sequence without a human clicking anything. The distinction from a regular chatbot with browsing is important: a chatbot with web search fetches a page and summarizes it. An agentic browser opens the page, clicks a button, fills out a shipping address, adds an item to a cart, and reports back once the task is done. All three products profiled here (Comet, Atlas, and Gemini Agent) sit on top of standard Chromium or WebView rendering engines, meaning the innovation is entirely in the automation and reasoning layer, not in a new way of rendering HTML.
The category exists because every major AI lab reached the same conclusion around the same time in 2025: chat interfaces cap out at answering questions, but a huge share of what people actually want done online (booking a flight, comparing five product listings, filling out a government form) requires taking action across multiple sites. OpenAI, Perplexity, and Google each built a version of that action layer. What’s different a year later is which of those bets paid off. Perplexity kept iterating and expanded to mobile. Google folded its standalone experiment into existing products. OpenAI is retiring its standalone browser entirely and redistributing the agent features into the core ChatGPT app and a Chrome extension instead, a move TechCrunch reported on July 9, 2026, framing it as OpenAI’s ambitions “still growing” even as the standalone app disappears.
Meet the Three Contenders
Perplexity Comet: The Only One Still Standing as a Standalone Browser
Comet is Perplexity’s own browser, first released in limited form in early 2025 and opened up for free worldwide access later that year. By August 2026 it ships on Mac, Windows, iOS, and Android, with the Android release landing on August 19, 2026, and marketed on the Google Play listing as “the first agentic AI browser built for mobile.” Perplexity’s product page describes Comet as a “Personal AI Assistant” built to act on a user’s behalf inside the browser itself, rather than requiring a separate app or extension. Deep Research inside Comet runs on a more capable reasoning model as of mid-2026, and Perplexity’s own release notes for Comet’s companion product, Comet Computer, describe Deep Research being extended with faster command access, task forking, inline action confirmations, and an analytics API aimed at enterprise customers managing usage limits.
Comet’s research pipeline for Max and Pro subscribers runs on Claude Opus 4.5, Anthropic’s flagship model tuned for long, source-heavy research work, giving Perplexity a reasoning backbone it doesn’t have to train in-house. Its distribution strategy also stands out: Comet’s agentic browsing was integrated into Samsung Internet in mid-2026, and Perplexity can be set as the default search engine on Samsung devices, giving it a built-in install base that neither Atlas nor Gemini Agent can claim through a single partnership.
ChatGPT Atlas: Built Fast, Retired Faster
OpenAI introduced Atlas in October 2025 as a Mac-first browser with ChatGPT built into the sidebar and an Agent Mode capable of navigating multiple webpages, filling out forms, logging into accounts the user has approved, and pulling data together to complete a stated goal like planning a trip or comparing products. BleepingComputer’s coverage of Atlas’s agent mode noted it could “open tabs and click through workflows for you, with safety limits and extra caution on sensitive sites.” Early 2026 updates added Saved Prompts, a DevTools-style Device Tab, tab renaming, and smoother handoffs between chat mode and agent mode, according to OpenAI’s own release notes archive.
Then, on August 22, 2026, OpenAI published a help center article titled “Evolving Atlas into ChatGPT for browser-based agentic work.” It states plainly that Atlas was scheduled to stop functioning as of August 9, 2026, and that after that date, “Atlas may no longer open, browse, or support browser-based agentic workflows.” Instead of a standalone browser, OpenAI is routing the same automation features through the main ChatGPT desktop app and a Chrome browser extension. TechCrunch’s July reporting had already flagged the shift was coming, describing OpenAI as “redistributing the agentic browsing features it tested in Atlas across ChatGPT’s desktop app and a Google Chrome extension” rather than abandoning the underlying technology.
Google Gemini Agent and Chrome Auto Browse: Project Mariner’s Second Life
Google never shipped a standalone AI browser in the way OpenAI and Perplexity did. Instead, it ran an experimental agent called Project Mariner starting in December 2024, capable of navigating Chrome, filling out forms, searching listings, and booking travel by taking screenshots and visually recognizing on-page elements. Google shut Mariner down as a standalone product on May 4, 2026, after a 17-month run, a move covered by outlets including Gagadget and AI Chat Daily, both of which reported that Mariner’s core navigation algorithms were absorbed into Gemini Agent rather than discarded.
What replaced Mariner is a two-part system. Gemini Agent, described in Google’s own materials as a “24/7 digital partner,” is available to Google AI Pro and Ultra subscribers and integrates directly with Gmail, Calendar, and Drive, plus a “require human review” toggle that forces a manual confirmation step before the agent takes a consequential action. Chrome Auto Browse, the second half, runs multi-step tasks embedded directly inside the Chrome browser itself rather than through a chat interface, and became available on desktop first, with an Android rollout that followed by late June 2026. Because Auto Browse lives inside Chrome, the world’s most-used browser, Google’s agentic reach is arguably the widest of the three, even without a dedicated “AI browser” brand name attached to it.
Perplexity Comet vs ChatGPT Atlas vs Gemini Agent: Full Specs Comparison
The table below lines up the three products across the specs that actually determine whether one fits a given workflow. Note that Atlas is included for reference even though it stopped functioning as a browser on August 9, 2026, since millions of users built workflows around it and need a clear picture of what changed.
| Spec | Perplexity Comet | ChatGPT Atlas | Gemini Agent + Chrome Auto Browse |
|---|---|---|---|
| Status as of Aug 31, 2026 | Active, expanding to Android | Discontinued Aug 9, 2026 | Active, folded into Chrome |
| Launch date | Early 2025 (limited), free worldwide late 2025 | October 2025 | Predecessor Mariner: Dec 2024; Agent/Auto Browse: 2026 |
| Platforms | Mac, Windows, iOS, Android | macOS only (Windows never shipped) | Chrome desktop; Android rollout by late June 2026 |
| Standalone app required | Yes, dedicated Comet browser | Yes (now retired) | No, built into existing Chrome |
| Underlying research model | Claude Opus 4.5 for Deep Research (Max/Pro) | Tuned OpenAI reasoning model in Agent Mode | Gemini 3.1 Pro and Gemini 2.5 Computer Use model |
| Context window | Depends on routed model; optimized for search, not long documents | Vendor-specific, not separately published for Atlas | Up to 1M tokens on Gemini 3.1 Pro |
| Multi-tab task automation | Yes, native to browsing engine | Yes, “open tabs and click through workflows” | Yes, via Chrome Auto Browse embedded layer |
| Form filling and checkout | Yes, agentic navigation and forms | Yes, shopping and trip-planning flows | Yes, forms, listings, and travel booking |
| Workspace/email integration | Limited, search-first design | ChatGPT connectors and connected apps | Deep: Gmail, Calendar, Drive via Gemini Agent |
| Human-review safeguard | Inline action confirmations (Comet Computer) | “Safety limits and extra caution on sensitive sites” | Explicit “require human review” toggle |
| Structured output generation | Presentations, spreadsheets, dashboards, websites | Text-based reports and completed web tasks | Workspace documents via Gmail/Calendar/Drive actions |
| Mobile agentic browsing | Yes, since Aug 19, 2026 (Android) | Never shipped on mobile | Auto Browse on Android since late June 2026 |
| Distribution partnership | Built into Samsung Internet, default search on Samsung | None announced | Native to Chrome, the world’s most-used browser |
Two things jump out from that table. First, Gemini Agent’s approach of living inside an existing browser rather than asking users to switch avoids the adoption friction that likely contributed to Atlas’s short lifespan. Second, Comet is the only product of the three that is both a dedicated standalone browser and expanding its platform footprint rather than shrinking it, which is why several independent reviewers, including a comparison published by IntuitionLabs, treat Comet as the baseline other agentic browsers get measured against.
Pricing Compared: What You Actually Pay for Agentic Browsing
None of the three companies sells its AI browser as a separate line item. Instead, agentic browsing rides on top of each company’s general-purpose subscription tiers, which makes direct price comparison slightly misleading unless you already pay for the underlying plan. Here’s how the tiers stack up as of August 2026.
| Tier | Perplexity | OpenAI (ChatGPT) | Google AI |
|---|---|---|---|
| Free tier | $0, Comet browser included | $0, browsing available; Atlas discontinued | $0, basic Gemini access |
| Entry paid tier | Pro: $20/mo ($16.67/mo billed annually) | Go: $8/mo | AI Pro: $19.99/mo |
| Mid paid tier | — | Plus: $20/mo | — |
| Top individual tier | Max: $200/mo | Pro: $200/mo | AI Ultra: cut from $250 to $200/mo |
| Team/enterprise seat | Enterprise Pro: $40/seat; Enterprise Max: $325/seat | Business/Enterprise: custom pricing | Workspace-linked AI add-ons: custom pricing |
| Agentic features gated behind | Free for basic Comet use; Deep Research/Comet Computer favors Pro/Max | Was bundled into all Atlas access, now shifting into Plus/Pro ChatGPT tiers | Gemini Agent gated behind AI Pro/Ultra subscription |
Perplexity’s pricing structure is the most permissive of the three for casual users: the core Comet browser and its everyday agentic browsing are free, with the paywall only kicking in around Deep Research depth and Comet Computer’s enterprise controls. Google keeps its best agentic features, Gemini Agent’s Workspace integration and Chrome Auto Browse’s more complex flows, locked behind the $19.99-a-month AI Pro tier or higher. OpenAI’s approach with Atlas had been the most generous in one sense (agent mode worked across free and paid ChatGPT tiers inside the app) but that entire access model disappears along with the app itself.
Benchmarks and Independent Test Results
Hard, numeric task-completion benchmarks for agentic browsers are still thin on the ground industry-wide, a gap several reviewers have flagged directly. What does exist is a consistent pattern of qualitative testing across independent outlets that lands on similar conclusions from different angles.
- IntuitionLabs’ technical review of Atlas evaluated task autonomy, stability, and safety qualitatively rather than with a numeric score, concluding that Agent Mode was “one of Atlas’s most talked-about features” but stopping short of publishing a completion-rate percentage, a gap that mirrors most independent AI browser testing to date.
- TechCrunch’s July 2026 reporting on the Atlas shutdown treated the agentic features tested in Atlas as a proving ground, reporting that OpenAI’s own internal read was that the experiment justified building more robust, integrated agentic tooling elsewhere rather than continuing to invest in a standalone app.
- A pre-Google I/O technical comparison from developer blog KSPL Academy assessed Gemini’s agent stack against rival browsing tools on capability depth rather than speed, highlighting the “require human review” toggle and Workspace integration (Gmail, Calendar, Drive) as differentiators for reliability and safety, not raw execution speed.
- A 2026 multi-chatbot comparison published by AI Insider tested ChatGPT, Claude, Gemini, and Perplexity side by side and concluded Perplexity was “best for cited, real-time research,” while naming Gemini Deep Research “the best no-cost alternative” for users who want long, structured, sourced answers without paying.
The absence of a standardized numeric benchmark (something like a shared “agentic browsing SWE-bench”) is itself a signal about how early this category still is. Every reviewer above independently reached for qualitative framing, safety controls, integration depth, citation quality, rather than a completion-rate percentage, because no accepted third-party test suite yet exists for autonomous browser tasks the way it does for coding models. Treat any article that claims a precise “87% task completion rate” for these products with skepticism unless it links to a named, reproducible test methodology.
Platform Reach and Distribution: Why Chrome’s Position Matters
Distribution turns out to be as important to this comparison as any feature checklist, because an agentic browser is only useful if people actually open it. None of the three companies has published hard install or monthly-active-user figures specific to the agent features themselves, so the table below compares distribution channels instead, which is the closest proxy available from public reporting.
| Distribution factor | Perplexity Comet | ChatGPT Atlas | Gemini Agent / Chrome Auto Browse |
|---|---|---|---|
| Requires a new browser install | Yes | Yes (now moot, discontinued) | No, ships inside Chrome |
| Pre-installed on any device | No, but bundled into Samsung Internet as of mid-2026 | No | Yes, wherever Chrome is already the default browser |
| Mobile OS coverage | iOS and Android | None (Mac-only, desktop app) | Android via Auto Browse; iOS coverage not separately confirmed |
| Default search engine option | Yes, on Samsung devices | No | Yes, Google is the default search engine on most devices already |
| User acquisition strategy | Free download plus OEM browser partnership | Bundled with existing ChatGPT subscriber base | Feature rollout to an installed base of over 3 billion Chrome users |
That last row explains a lot about why Google’s approach survived where OpenAI’s didn’t, even though Google killed its own original experiment, Project Mariner, before OpenAI killed Atlas. Chrome’s existing install base meant Google never needed anyone to download anything new; it just needed to ship a feature update. Perplexity took the opposite bet and it’s paying off differently: rather than fight for browser market share head-on, Comet is riding into Samsung’s default browser slot on a partnership basis, which functions like a distribution shortcut without requiring Comet itself to become the world’s most popular browser outright. OpenAI’s Atlas had neither advantage. It asked users to abandon their existing browser for a brand-new one built by a company with no browser market share to leverage, and the ten-month lifespan suggests that ask didn’t scale the way a Chrome feature update or an OEM partnership could.
Agentic Capabilities: Form Filling, Checkout, and Multi-Tab Workflows
All three products can do the same broad category of task: read a page, decide what to click, fill in a field, move to the next page, and repeat until a goal is met. The differences show up in scope and guardrails rather than raw capability.
Atlas’s Agent Mode, per OpenAI’s own product page, gave ChatGPT what the company described as a “virtual computer” inside its sandbox, letting it navigate multiple webpages, fill out forms, log into accounts a user had approved, and pull data together toward a stated goal such as shopping or trip planning. BleepingComputer’s testing found it could open tabs and click through workflows with what OpenAI called “extra caution on sensitive sites,” a soft signal that the model deliberately slows down or asks for confirmation on things like payment forms.
Gemini Agent’s equivalent capability leans on the DOM-reading and multi-step planning engine inherited from Project Mariner, which AI Wiki describes as able to “read the DOM, plan multi-step browser flows, and click through forms.” Reporting on Mariner’s original capabilities found it could navigate Chrome, fill out forms, search listings, and book travel by taking screenshots and visually recognizing page elements, a computer-vision-adjacent approach rather than pure DOM parsing, which likely explains why the technology generalized well enough to survive the Mariner shutdown and get folded into two separate products instead of being killed outright.
Comet’s agentic browsing, per Perplexity’s own materials, is framed around research and task execution together rather than pure automation: the browser is built to act on stated intent rather than literal instructions. Comet Computer’s 2026 update added the ability to fork a running task into parallel branches and confirmed actions inline before execution, giving users a checkpoint system that neither Atlas nor Gemini Agent has publicly matched at the same granularity.
5 Real-World Use Cases Where Agentic Browsers Actually Get Used
The marketing copy for all three products leans on abstractions like “autonomous task completion.” Here’s what that looks like in practice, based on the documented capabilities above.
- Comparison shopping across multiple retailer sites. A user states a product and a budget; the agent opens several retailer tabs, compares prices and shipping estimates, and either completes checkout or returns a shortlist depending on how much autonomy the user granted. This is the scenario OpenAI cited most often for Atlas’s Agent Mode, and it’s also the flow Gemini Agent’s form-filling and multi-tab planning were originally built around inside Project Mariner.
- Trip planning and travel booking. Reporting on Project Mariner’s original capabilities specifically called out that it could “search listings, and book travel,” a workflow that carried over directly into Gemini Agent. Atlas offered the same category of task before its shutdown.
- Cited, source-heavy research reports. This is Comet and Perplexity’s core strength, and it’s the use case independent reviewers consistently rank Perplexity highest on, since every claim in a Comet-generated report links back to a source rather than presenting a synthesized answer with no citation trail.
- Inbox and calendar triage. Gemini Agent’s direct integration with Gmail, Calendar, and Drive lets it draft replies, schedule meetings, or pull a document referenced in an email thread without the user switching apps, a workflow neither Comet nor Atlas can match without third-party connectors.
- Generating a deliverable, not just an answer. Comet Computer’s 2026 update lets Deep Research output presentations, spreadsheets, dashboards, or even a simple website directly, moving past the text-report format that both Atlas and Gemini Agent still rely on for most research tasks.
Security and Prompt Injection Risks in Agentic Browsers
Giving a language model the ability to click buttons and fill in forms on your behalf raises the stakes on a well-known weakness: prompt injection, where malicious text hidden on a webpage tricks the model into taking an action the user never asked for. None of the sources reviewed for this comparison document a confirmed, named 2026 security incident where Comet, Atlas, or Gemini Agent was successfully exploited in the wild. That absence is worth treating carefully rather than as reassurance, since it may reflect limited public disclosure as much as an absence of attempts.
What each vendor has published are mitigation features rather than incident reports. OpenAI built “safety limits and extra caution on sensitive sites” into Atlas’s Agent Mode, according to BleepingComputer’s testing. Google shipped an explicit “require human review” toggle inside Gemini Agent that forces manual confirmation before a consequential action executes. Perplexity’s Comet Computer update added inline action confirmations as a checkpoint before a task proceeds. All three approaches boil down to the same idea: insert a human checkpoint before the agent does something irreversible, like submitting a payment or sending an email.
For site owners worried about agentic browsers scraping or interacting with their pages in unintended ways, the practical defense mirrors how sites already handle aggressive crawlers: a robots.txt directive targeting known AI agent user agents, combined with rate limiting on form submission endpoints.
# Example robots.txt entries some site operators use
# to signal AI browsing agents should not auto-navigate forms
User-agent: ChatGPT-User
Disallow: /checkout/
Disallow: /account/
User-agent: PerplexityBot
Disallow: /checkout/
Disallow: /account/
User-agent: Google-Extended
Disallow: /checkout/
Disallow: /account/Readers building or securing AI-facing applications more broadly should also look at how prompt injection defenses apply beyond the browser layer. Our guide to preventing prompt injection attacks covers the same threat model in the context of chatbot apps and AI agents generally, and most of the mitigations there (input sanitization, allow-listing trusted domains, sandboxing high-risk actions) apply directly to agentic browsers too.
Why ChatGPT Atlas Is Being Discontinued
OpenAI has not published a single blunt explanation for retiring Atlas, but the pattern across its own release notes and independent reporting points to a consolidation strategy rather than a failure of the underlying technology. The help center article announcing the shutdown frames it as “evolving Atlas into ChatGPT for browser-based agentic work,” language that treats the standalone browser as a stepping stone rather than a dead end. TechCrunch’s July reporting reached a similar conclusion, noting OpenAI is “redistributing” the agentic features tested in Atlas “across ChatGPT’s desktop app and a Google Chrome extension” instead of scrapping them outright.
Two structural problems likely drove that decision. First, Atlas launched Mac-only and never shipped a Windows version in its roughly ten months of life, capping its addressable audience against Comet’s four-platform footprint and Chrome’s built-in reach through Auto Browse. Second, asking users to install and switch to an entirely new browser is a much higher-friction ask than adding agent capabilities to a browser extension or an app people already have open, which is exactly the model Google used with Chrome and the one OpenAI is now adopting.
Migration Guide: Moving Off ChatGPT Atlas
If you built bookmarks, saved prompts, or automated workflows inside Atlas before its August 9, 2026 cutoff, here’s how to move that work forward without losing continuity.
- Export or note down any Saved Prompts you built in Atlas. These were stored inside the Atlas app itself and won’t transfer automatically. Copy the text into a notes app before you lose access.
- Check whether your Atlas workflows already work in the ChatGPT desktop app. OpenAI’s stated plan is to move Agent Mode into ChatGPT’s existing desktop client, so re-test your most common tasks there first.
- Install the ChatGPT Chrome extension as the direct successor to browsing inside Atlas itself, since OpenAI is routing browser-based agentic work through this channel going forward.
- Re-authenticate any connected accounts (email, calendar, shopping accounts) that Atlas had permission to access, since those authorizations don’t carry over to a new client automatically.
- If your workflow leaned heavily on multi-tab automation, trial Perplexity Comet in parallel. It’s the most direct like-for-like replacement since it’s also a standalone browser rather than an extension bolted onto an existing one.
- If your workflow leaned on Gmail, Calendar, or Drive automation, Gemini Agent is the stronger fit given its native Workspace integration, though it requires a Google AI Pro or Ultra subscription.
- Rebuild any Device Tab or DevTools-style debugging routines you relied on in Atlas using your browser’s native developer tools, since that feature was Atlas-specific.
- Test checkout and payment flows manually before trusting an agent with them again in whichever new tool you land on, since permission scopes and safety guardrails differ across products.
- Set up the “require human review” toggle if moving to Gemini Agent, or the equivalent inline confirmation in Comet Computer, to avoid surprises during the transition period while you learn the new tool’s behavior.
- Audit which sites you’d previously granted Atlas login access to and revoke any that you no longer need automated, cleaning up your permission footprint as part of the move.
- Compare Deep Research quality directly by running the same research query through Comet and Gemini Agent side by side before committing to one as your default.
- Budget for a subscription change if your prior workflow assumed Atlas’s free-with-ChatGPT access, since Gemini Agent’s best features sit behind the $19.99/mo AI Pro tier and Comet’s Deep Research depth favors Perplexity’s $20/mo Pro plan.
Pros and Cons of Each AI Browser
Perplexity Comet
Pros: Available free across four platforms including Android as of August 19, 2026; strongest citation and source-transparency track record among the three; Claude Opus 4.5 backing for Deep Research; built-in Samsung Internet distribution; structured output generation (spreadsheets, dashboards, sites) beyond plain text reports.
Cons: Weaker native Workspace-style integration than Gemini Agent; requires installing and switching to a dedicated browser rather than extending one you already use; enterprise-grade controls like the analytics API sit behind Comet Computer, a separate paid layer.
ChatGPT Atlas (Historical, for reference)
Pros (while active): Deep integration with ChatGPT’s existing memory and Projects features; broad Agent Mode capability across shopping, trip planning, and general web tasks; safety-conscious design with extra caution flagged on sensitive sites.
Cons: Discontinued as of August 9, 2026; never shipped on Windows in its roughly ten-month life; users must now migrate workflows to the ChatGPT desktop app or a Chrome extension instead.
Google Gemini Agent + Chrome Auto Browse
Pros: No new browser to install, since it runs inside Chrome, the world’s most widely used browser; deepest Workspace integration (Gmail, Calendar, Drive) of the three; explicit “require human review” safety toggle; up to 1M-token context on Gemini 3.1 Pro for long documents.
Cons: Best features gated behind the $19.99/mo AI Pro tier or the $200/mo Ultra tier; Android rollout for Auto Browse lagged desktop by roughly a month; less citation-forward than Comet for pure research tasks.
Which AI Browser Should You Use? 5 Recommendations by Scenario
- If you’re a heavy researcher who needs every claim cited: choose Perplexity Comet. Its source-transparency track record is the strongest of the three, and Deep Research is now backed by Claude Opus 4.5 for Max and Pro subscribers.
- If your work lives inside Gmail, Calendar, and Drive: choose Gemini Agent. Its native Workspace integration has no equivalent in Comet or the post-Atlas ChatGPT tooling.
- If you were an Atlas power user with existing ChatGPT Projects and memory: stay inside the ChatGPT ecosystem and move to the desktop app plus the Chrome extension, since your existing context and history carry over most cleanly there.
- If you need agentic browsing on Android specifically: Comet is currently the most mature mobile option, having shipped a dedicated Android agentic browser on August 19, 2026, ahead of a full-featured Auto Browse equivalent on that platform.
- If you manage a team and need usage controls and an audit trail: Comet Computer’s analytics API and custom credit limits, aimed squarely at enterprise administrators, currently offer more granular usage governance than the consumer-facing Gemini Agent toggle.
The Verdict: Comet Wins on Survival, Gemini Wins on Reach
Judged purely on which product is still standing and actively expanding, Perplexity Comet wins this comparison by default: it’s the only one of the three that remains a dedicated, growing, cross-platform AI browser as of August 31, 2026. But “still exists” isn’t the same as “best for every job.” Gemini Agent’s decision to live inside Chrome rather than ask for a new install gives it the broadest practical reach, and its Workspace integration covers a category of daily task (email, calendar, documents) that neither Comet nor the post-Atlas ChatGPT tooling matches natively. OpenAI’s retreat from a standalone browser, meanwhile, looks less like a loss for the underlying agent technology and more like an admission that the browser wrapper itself was the wrong distribution bet, a lesson Google seems to have learned a year earlier when it kept Mariner’s tech inside Chrome from the start.
For most individual users, the pragmatic move is to try Comet first, since it’s free, cross-platform, and requires no existing subscription to test. Power users already paying for Google AI Pro or Ultra should give Gemini Agent a real trial before switching services, since the Workspace integration alone can offset its higher price floor. And anyone still relying on Atlas needs to move now: the app stopped functioning as of August 9, 2026, and every workflow built inside it needs a new home.
What Comes Next for Agentic Browsing
The pattern across 2026 suggests the AI browser category is consolidating around two models rather than three: a dedicated standalone browser (Comet’s approach) and an agent layer bolted onto an existing browser or app people already use daily (Google’s Chrome-first approach, and now OpenAI’s post-Atlas pivot). A truly standalone third browser competing against both looks like a harder sell going forward, given how quickly Atlas’s ten-month run ended. Expect the next round of competition to focus less on “who has a browser” and more on which agent handles sensitive actions, payments, account logins, personal data, most safely, since that’s the dimension every vendor is now racing to publish new safeguards around rather than new browsing features.
For a broader look at how these companies’ underlying chat products stack up outside the browser context, see our comparisons of Claude vs Perplexity, Perplexity vs ChatGPT vs Gemini, and ChatGPT vs Gemini vs Claude Pro, plus our look at how Copilot, Gemini, and Perplexity compare on enterprise pricing and user scale.
Frequently Asked Questions
What happened to ChatGPT Atlas?
OpenAI scheduled Atlas to stop functioning as a browser on August 9, 2026. According to OpenAI’s help center article published August 22, 2026, the agentic browsing features tested in Atlas are being moved into the main ChatGPT desktop app and a Chrome extension rather than continuing as a standalone product.
Is Perplexity Comet free to use?
Yes. Comet’s core browser and everyday agentic browsing are free across Mac, Windows, iOS, and Android. Deeper Deep Research capability and Comet Computer’s enterprise controls are tied to Perplexity’s paid Pro ($20/mo) and Max ($200/mo) tiers.
What is Google’s Gemini Agent, and how is it different from Project Mariner?
Gemini Agent is Google’s consumer-facing agentic assistant, built on the navigation technology from Project Mariner, an experimental browser agent Google shut down as a standalone product on May 4, 2026. Mariner’s core algorithms now power Gemini Agent’s task automation and Chrome Auto Browse’s in-browser task completion.
Can these AI browsers actually complete an online checkout on their own?
All three have documented shopping and checkout-style workflows. Atlas’s Agent Mode explicitly supported shopping and trip planning before its shutdown. Gemini Agent inherited form-filling and travel-booking capability from Project Mariner. Comet supports agentic navigation and forms as part of its core browsing engine. All three vendors describe extra caution or explicit human-review steps around payment and account actions specifically.
Are agentic AI browsers safe from prompt injection attacks?
No vendor claims complete immunity. Each has published mitigation features rather than a guarantee: OpenAI built safety limits and extra caution into Atlas’s Agent Mode, Google added a “require human review” toggle to Gemini Agent, and Perplexity added inline action confirmations to Comet Computer. None of the sources reviewed here document a confirmed, publicly named 2026 exploitation incident against any of the three products, but that shouldn’t be read as proof the risk doesn’t exist.
Which AI browser is best for research with citations?
Perplexity Comet, according to independent 2026 testing that specifically ranked Perplexity as “best for cited, real-time research,” with Gemini Deep Research noted as the strongest free alternative for users who want long, structured, sourced answers without a paid subscription.
Do I need a paid subscription to use agent features on any of these?
Comet’s baseline agentic browsing is free. Gemini Agent’s fullest capabilities are gated behind Google’s AI Pro ($19.99/mo) or AI Ultra ($200/mo) tiers. ChatGPT’s post-Atlas agentic tooling inside the desktop app and Chrome extension follows ChatGPT’s own tier structure, ranging from a $8/mo Go plan up to $200/mo Pro.
What should I use instead of ChatGPT Atlas now that it’s shut down?
OpenAI’s own recommendation is to move to the ChatGPT desktop app combined with its Chrome extension, since that’s where the company is routing Atlas’s former agentic features. Users who want a genuinely standalone browser experience similar to Atlas’s original design should trial Perplexity Comet, which remains the closest like-for-like replacement still actively developed as of August 2026.