AI Search
We tested 10 AI browsers across 4 categories to answer: Do these actually make browsing better, or are they chatbots glued onto Chrome?
The short answer: Most are somewhere in between. A few deliver genuine productivity gains. Others promise a lot and fail during actual use.
As a result of our testing, AI browsers show significant accessibility barriers and performance gaps that limit practical adoption for most users.
Task: “Summarize AIMultiple’s main page”
We asked each browser to summarize AIMultiple’s homepage and a long technical article on agentic AI for cybersecurity.
Perplexity Comet: Navigated to the site independently, analyzed content, and delivered structured summaries with specific examples.
ChatGPT Atlas: Analyzed pages through its sidebar. When browser memory is enabled, it connects current content to previous browsing history. You can ask follow-up questions about specific sections without re-explaining the context.
Microsoft Edge Copilot: Correctly identified key sections, AI benchmarks, LLM calculators, and enterprise software insights. Solid understanding of business content.
Brave Leo: Accurately covered enterprise software insights, AI benchmarks, calculators, and the site’s transparency focus. Well-structured response.
Arc Max: Can’t perform standalone summarization. AI functionality only activates when you right-click specific page elements. There’s no chat interface where you can ask, “Summarize this page.”
Opera Aria: Failed both tests. Instead of analyzing actual page content, it provided generic LLM responses. The page context feature appears broken; it seems to be unable to see what you’re looking at.
Sigma AI: Unable to access external websites directly. The browser explicitly states it cannot visit URLs and requires manual text input to provide summaries, severely limiting web summarization capabilities.
Strawberry Browser: Still in alpha. Early demos suggest strong autonomous capabilities, but a comprehensive evaluation isn’t possible yet due to limited access.
Analyze article: Agentic AI for Cybersecurity: Real life Use Cases & Examples
Perplexity Comet: Delivered structured analysis covering SecOps and AppSec use cases, cited specific examples (University of Kansas Health System, APi Group), and broke down benefits and challenges.
Microsoft Edge Copilot: Organized findings into clear sections: SecOps automation, AppSec uses, and implementation challenges.
Brave Leo: Covered autonomous AI operations, SecOps/AppSec applications, automation benefits, and challenges. Strong grasp of technical concepts, suggested follow-up questions.
ChatGPT Atlas: Analyzed the article with context awareness, breaking down technical concepts and offering to compare with similar articles from browser memory.
Arc Max: Provided detailed analysis across multiple attempts but was repetitive. Captured key concepts like autonomous decision-making, real-time monitoring, and SOC automation, though less concise than competitors.
Opera Aria: Context functionality broken. Defaulted to generic responses instead of analyzing the actual article.
We tested each browser between August 2024 and January 2025 as they became available:
What we tested:
These browsers split into two camps:
Smart assistants: Add AI chat and analysis, but you still control the browsing. (Arc Max, Brave Leo, Microsoft Edge Copilot, ChatGPT Atlas sidebar)
AI agents: Browse autonomously, make decisions, and complete tasks without constant guidance. (Perplexity Comet agent, ChatGPT Atlas agent mode, Strawberry Browser)
Whether you need an agent depends on your workflow. If you’re researching a topic across dozens of sites, an agent saves hours. If you just want quick summaries while reading, a smart assistant is enough.
Atlas brings ChatGPT into your browser as a core feature, not an extension.
Core features:
Agent Mode (Plus/Pro subscribers only):
ChatGPT navigates websites, fills forms, books reservations, and adds items to carts, all with your permission, before important actions.
Example workflow: “Find Italian restaurants in downtown Seattle with availability Saturday night, and book a table for 4 at the one with the best reviews.”
Atlas opens multiple tabs, reads reviews, checks availability, and presents options. You approve the booking, and the reservation is completed.
Safety limits: Can’t run code, download files, install extensions, access other apps, read passwords, or use autofill data.
Search & Browse:
ChatGPT search opens with an AI-generated response, then provides tabs for traditional results, images, videos, and news. Live results pull from current web data. Maintains context across multiple tabs and websites.
Privacy & Data:
Browser memories are opt-in. Incognito mode is available browse, signed out, with no chat or memory saved.
By default, browsing content isn’t used for training. You can opt in via data controls.
Chat history is stored according to your ChatGPT account settings. Data retained 30 days then deleted.
Integration:
Currently, macOS only. Windows, iOS, Android coming (no date specified).
Incentive program: Increased ChatGPT usage limits if you keep Atlas as the default browser for 7 consecutive days.
The catch: Agent mode is the killer feature, but it requires a paid subscription. Free users get a smart assistant, not an autonomous agent. And it’s Mac-only for now; Windows users are waiting.
Comet demonstrates true autonomous browsing. Give it multi-step tasks, and it figures out how to complete them.
Comet Assistant:
Cross-tab context maintains conversation across different websites. Screen awareness AI sees what you’re looking at without screenshots or copy-paste.
Can book restaurants, make purchases, schedule meetings autonomously. “Find the cheapest direct flight to Tokyo departing next Tuesday” triggers an actual search across multiple travel sites.
Search Integration:
Perplexity AI search is pre-installed as the default. AI-generated summaries with source citations. Sidebar accessible on all websites for contextual queries.
Mobile & Background features:
Mobile app in preview designed specifically for phone interfaces with voice technology.
Background assistant (Max users) can perform multiple tasks simultaneously. Works while you’re away from the computer.
The catch: Security researchers at Brave discovered a vulnerability that allows attackers to embed hidden instructions in web content. When you ask Comet to summarize a page, it might execute malicious commands without distinguishing them from legitimate content.
Perplexity claims they’ve added mitigations, but this is an inherent risk with autonomous AI agents. They can be tricked through carefully crafted webpage content.
Arc takes a different approach, no chat interface, just context-specific AI features.
AI capabilities:
“Ask on Page”: Use Command-F/Control-F to ask questions about webpage content instead of just searching for keywords.
“5-Second Previews”: Hover + Shift over links for instant webpage summaries (macOS only).
Browser automation & organization:
“Tidy Tab Titles”: Automatically rename pinned tabs with shorter, clearer titles.
“Tidy Downloads”: Smart file renaming based on content and context.
“Tidy Tabs”: Auto-organize tabs when you have more than six open.
Data sharing:
Max features send data to AI partners (OpenAI) with clear disclosure. The “Browse for Me” feature uses zero data retention (ZDR), which processes but does not store data.
The catch: No traditional chat interface. You can’t ask “What’s the main point of this article?” unless you trigger the specific Command-F feature. The AI is powerful but locked behind specific interaction patterns that aren’t intuitive.
Built into the Edge browser with Microsoft ecosystem integration.
Copilot features:
Native AI sidebar in browser. Voice-activated interactions. Page analysis provides insights into current content. Content generation for emails, documents, and summaries within the browser.
Microsoft ecosystem:
Office 365 connectivity integrates with Word, Excel, PowerPoint, Outlook. Microsoft Graph access to calendar, emails, and documents across services. Teams integration for meeting scheduling. OneDrive for document access and sharing.
Productivity features:
Shopping assistant tracks prices, finds coupons, compares products. Travel planning with itinerary creation and booking assistance. Reading mode enhanced with AI summarization. Real-time webpage translation.
Search & Browse:
Bing Chat integration: AI-powered search with a conversation interface. Web Compose generates content directly in web forms. Instant webpage and PDF summarization.
The catch: Edge Copilot works best if you’re paying for Microsoft 365. Without it, you get basic chat features but miss the ecosystem integrations that make Copilot actually useful. The free version feels incomplete.
Leo prioritizes privacy over advanced features. No signup, no payment, no data collection.
Summarizes webpages, PDFs, and YouTube videos. Answers questions about current page. Coding assistance. Voice input on mobile. Real-time search integration for current information.
Privacy design:
No account required for the free version. Conversations are stored locally on the device, not cloud servers. Responses discarded after generation are not used for training.
Anonymous usage requests cannot be linked to users. No IP logging. No server-side records retained.
Model choices:
Free tier: Llama, Qwen, Gemma models. Premium ($14.99/month): Claude Sonnet 4, higher rate limits.
Bring Your Own Model (BYOM): Connect local models via Ollama or third-party APIs like GPT-4. Use your own keys.
Browser integration:
Built into the Brave sidebar, address bar, and full-page mode. Cross-platform Windows, macOS, Linux, Android, iOS.
Page context awareness works with webpages, PDFs, and YouTube videos. Multi-tab context for working across multiple tabs (advanced feature).
The catch: Privacy comes with trade-offs. Free models are less capable than ChatGPT or Claude. Can’t perform autonomous actions, such as Comet or Atlas agent mode. Great for private research and summarization. Not for complex automation.
Opera promises the most 150+ local AI models, built-in model management, and image generation. In practice, core functionality doesn’t work.
150+ local LLM variants from 50 model families. Download and delete models directly in the browser. 2-10 GB storage per model. Supported models include Llama, Phi-2, Gemma, Vicuna, and Mixtral.
Image generation using Google’s Imagen. Image analysis with OCR and visual understanding (upload up to 3 images).
Integration:
Built into the Opera sidebar and address bar. Cross-platform support. Page context awareness. Quick access via Ctrl+/ or Cmd+/.
The catch: The page context, the most important feature, is broken. During our tests, Opera Aria consistently failed to analyze actual webpage content. Instead of summarizing the article you’re reading, it provides generic responses based on your query keywords.
This makes it nearly useless as a browser assistant. The local models work fine for general chat, but if it can’t see what you’re actually browsing, what’s the point?
Sigma has an AI chat interface but can’t access external websites.
AI Chat with multiple conversation modes. Image generation. Compose feature for content creation. SEO-friendly content for blogs and marketing.
Model Selection & Flexibility
Integration:
Pre-installed SigmaGPT extension. Cross-platform support. Dual search mode: traditional search or AI conversations. Built-in crypto wallet.
The catch: The browser explicitly states it cannot visit URLs. You have to manually input text to get summaries. For a “browser” assistant, this is a fatal limitation. It’s a chatbot that happens to be in a browser, not an AI browser.
AI Assistant Capabilities
Integration & Features
Alpha-stage browser focused on workflow automation.
AI capabilities:
AI companions work across websites autonomously. Screen recording learning AI observes your actions and learns. Cross-website automation behind login pages. Approval-based actions before execution.
Advanced features:
Multi-site research gathers information from hundreds of websites simultaneously. Data organization automatically compiles research into spreadsheets. Content generation learns your writing style and voice. Form automation handles repetitive data entry.
The catch: It’s alpha software. Expect bugs, limited functionality, and frequent changes. $30/month pricing positions it mid-range, but paying for alpha software is risky. Wait until it’s more mature unless you’re willing to beta test.
Google Disco is an experimental AI browser developed by Google Labs that combines traditional web browsing with AI-powered workspace features.
AI capabilities:
Integrated Gemini AI for conversational assistance. Workspace integration connection with Google Docs, Sheets, and Drive. Multi-modal search combining text, image, and voice. Smart Tabs automatically organize and group related tabs.
Core features:
Contextual assistance understands content across multiple tabs. Content summarization for articles, documents, and webpages. Research mode gathers information from multiple sources and compiles findings.
Google ecosystem integration:
Gmail, Calendar, and Meet integration. Google Drive access for document collaboration. Chrome extension compatibility. Cross-device sync across Android, ChromeOS, and desktop.
The catch: Its experimental access is limited, features change frequently, and there’s no guarantee Google won’t kill it as they do with most experiments.
Security researchers at Brave discovered a significant security flaw in Perplexity Comet’s implementation that allows attackers to manipulate AI actions through malicious web content1 .
Attack mechanism:
High Risk – Advanced Agentic Browsers:
Medium Risk – Limited Agentic Features:
Lower Risk – Assistant-Only Browsers:
Core AI Capabilities
Common Privacy Features
Standard Integration
Your email address will not be published. All fields are required.
