We Use Cookies

    We use cookies to enhance your browsing experience, analyze site traffic, and personalize content. By clicking "Accept All", you consent to our use of cookies. You can customize your preferences or reject non-essential cookies.

    Learn more about our cookie policy
    Feature Guide

    Three Ways to Search the Web with AI

    Compare Perplexity's built-in search, Native Agentic Search, and Exa Plugin—and choose the best for your needs.

    Each system has unique strengths: Perplexity for speed, Native for deep research, Exa for universal compatibility.

    Web Search
    8 min read
    Deep Dive

    Quick Summary

    We offer three complementary web search systems that work together to give you the best search experience:

    Perplexity Search

    Specialized search models with built-in web access. Optimized for speed and accuracy.

    ✅ Sonar, Sonar Pro, Deep Research

    Native Agentic Search

    Models with function calling autonomously search and fetch URLs. Full control, deep research.

    ✅ GPT-5, Claude 4.5+, Gemini 3+

    Exa Plugin Search

    AI-optimized semantic search for ANY model via OpenRouter plugin. Works with all 200+ models.

    ✅ DeepSeek, Qwen, Llama, older models

    1. Perplexity Native Search (Built-in)

    What it is: Perplexity models have search built into the model itself by the provider. No plugins needed, no toggles—it just works.

    🎯 How It Works:

    • 1. You select a Perplexity model (Sonar, Sonar Pro, etc.)
    • 2. Model automatically searches when needed (built-in capability)
    • 3. Search is optimized and integrated by Perplexity
    • 4. Results come back with citations and sources

    🔍 Perplexity Model Comparison:

    ModelBest ForSpeedQualityCostDepth
    SonarQuick Q&A, daily news, simple facts⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐
    Sonar ProBalanced research, most queries⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐
    Sonar Pro SearchComplex search queries, agentic use⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐
    Deep ResearchComprehensive analysis, reports⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐
    Reasoning ProComplex problem-solving + research⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐

    💡 Quick Pick:

    Start with Sonar Pro Search (highlighted) for best balance of speed, quality, and depth. Use Sonar for quick news, Deep Research for comprehensive analysis.

    Advantages:

    • • Fastest search: Optimized end-to-end (3-5 seconds)
    • • No configuration needed: Always works, no toggles
    • • Integrated citations: Sources built into response
    • • Optimized for search: Models trained specifically for web research
    • • Range of options: 5 models from fast (Sonar) to deep (Deep Research)

    Limitations:

    • • Perplexity-only: Only works with Perplexity models
    • • Cost included in model: Can't separate search from generation cost
    • • Less control: Can't control number of results or search depth

    💰 Pricing:

    • • Sonar: ~25 credits per query (search included)
    • • Sonar Pro/Pro Search: ~30 credits (search included)
    • • Deep Research/Reasoning Pro: 50-100 credits (multiple searches)

    Best for: Fast news checks, fact-checking, general research where speed matters, users who want simplicity (no configuration).

    2. Native Agentic Search (Function Calling)

    How it activates: When the Web Search toggle is OFF (or no preference set), models with native function calling can autonomously use search tools.

    🎯 How It Works:

    1. 1. Model decides it needs current information
    2. 2. Calls search_web("query", 10) → Gets 10 results from Brave Search
    3. 3. Reviews snippets and decides if it needs more detail
    4. 4. Calls fetch_url("specific-url") → Reads full page (up to 50k chars)
    5. 5. Can search again with refined queries
    6. 6. Synthesizes information from multiple sources

    Advantages:

    • • Full autonomy: Model controls entire research process
    • • Deep research: Can search → read → search again in loops
    • • Complete context: Reads full pages (50k chars), not just snippets
    • • Better for complex topics: Multi-step reasoning shines
    • • More results: 10 search results vs 5 (Exa)
    • • Follow discovery chains: Can find and fetch linked resources

    Limitations:

    • • Requires capable models: Only GPT-5+, Claude 4.5+, Gemini 3+
    • • Uses context window: Full pages consume model context
    • • Multiple tool calls: Each search/fetch billed separately
    • • Dependent on model quality: Small models struggle

    Available Web Search Tools:

    • Search Web - AI searches the web and finds relevant results
    • Fetch URL - AI reads full webpage content for deep analysis

    💰 Pricing:

    • • Search Web: 20 credits per search (10 results)
    • • Fetch URL: 1 credit per webpage
    • • Example: 2 searches + 3 URLs = 20+20+1+1+1 = 43 credits

    Best for: Research questions, fact-checking, technical documentation analysis, anything requiring deep investigation.

    3. Exa Plugin Search (OpenRouter)

    How it activates: When the Web Search toggle is ON for a model, we inject the Exa plugin via OpenRouter.

    🎯 How It Works:

    1. 1. You enable Web Search toggle for a model in settings
    2. 2. Our system injects Exa plugin into the request
    3. 3. Our system also blocks native tools (prevents conflicts)
    4. 4. Model decides when to search based on your question
    5. 5. OpenRouter calls Exa AI for semantic search
    6. 6. Model receives 5 results and generates answer

    Advantages:

    • • Works with ANY model: All 200+ models supported
    • • Semantic search: AI-optimized Exa engine
    • • Simpler for users: Just toggle on, model handles rest
    • • Per-model control: Enable for specific models only
    • • Single billing: One charge includes search
    • • No function calling needed: Works with older/smaller models

    Limitations:

    • • Fewer results: 5 results vs 10 (Native)
    • • No URL fetching: Only search snippets, no full pages
    • • No iterative research: One search per question
    • • Blocks native tools: Can't use search_web if toggle ON

    💰 Pricing:

    • • Exa search: ~20 credits per query (5 results)
    • • Included in generation: Billed as part of model response
    • • Example: DeepSeek R1 (~32 credits) + Exa search (20 credits) = 52 credits total

    Best for: Quick news checks, using non-AGENTIC models with search, consistent search behavior across all models.

    🎛️ The Web Search Toggle Explained

    The Web Search toggle controls which system is active:

    Toggle OFF (Default for AGENTIC models):

    ✅ Native tools ENABLED: Model can use search_web() and fetch_url()

    ❌ Exa plugin DISABLED: No OpenRouter plugin injected

    Result: Full agentic research with autonomy

    Toggle ON (Enable Exa):

    ✅ Exa plugin ENABLED: OpenRouter injects Exa semantic search

    ❌ Native tools BLOCKED: System prompt tells model "Don't use search_web()"

    Result: Single Exa search, works with all models

    ⚡ Key Insight:

    The toggle overwrites the native behavior for AGENTIC models. When ON, even capable models (GPT-5, Claude) will use Exa instead of native tools. This prevents conflicts—only ONE search system is active at a time.

    🎯 Decision Guide: Which Search Should I Use?

    Choose Native Agentic Search When:

    • You have an AGENTIC model (GPT-5, Claude 4.5+, Gemini 3+)
    • You need deep research with multiple sources
    • You want the model to read full articles, not just snippets
    • Complex questions requiring iterative research
    • You want maximum autonomy and model intelligence

    Example: "Research the latest quantum computing breakthroughs and explain the technical details"

    Choose Exa Plugin Search When:

    • You're using a non-AGENTIC model (DeepSeek, Qwen, Llama)
    • You need quick news or simple fact-checking
    • You want consistent search across all your models
    • Snippets are sufficient (don't need full pages)
    • You prefer predictable behavior (no autonomous decisions)

    Example: "What's today's top AI news?" or "Summarize recent developments in [topic]"

    🔧 Technical Details

    Search Engine Details

    Native Search:

    • • Engine: Brave Search API
    • • Results: 10 per query
    • • Type: Traditional web index
    • • Format: Titles, links, snippets

    Exa Plugin:

    • • Engine: Exa AI (semantic)
    • • Results: 5 per query
    • • Type: AI-optimized search
    • • Format: Contextual snippets

    Pricing Comparison (Same Cost)

    MethodResultsCostValue
    Native (search_web)10 results20 credits⭐⭐ Better
    Exa plugin5 results~20 credits⭐ Good

    💡 Same price, but native gives you 2× results plus URL fetching capability!

    📚 Real-World Examples

    Native Agentic
    Deep Technical Research

    You:

    "How does React's new compiler work? Include technical details."

    GPT-5 (Toggle OFF):

    [Calls: search_web("React compiler technical details 2026", 10)]
    [Reviews 10 results, finds React docs URL]
    [Calls: fetch_url("https://react.dev/compiler")]
    [Reads 15k words of technical documentation]
    [Searches for examples: search_web("React compiler code examples")]
    [Fetches example repository]

    "The React Compiler transforms your components at build time... [detailed technical explanation with code examples and citations from official docs]"

    💰 Cost: ~25 credits (model) + 20+20 (2 searches) + 1+1 (2 fetches) = 67 credits

    Exa Plugin
    Quick News Summary

    You:

    "What's the latest AI news?"

    DeepSeek R1 (Toggle ON):

    [Exa plugin searches automatically]
    [Receives 5 semantic results from Exa AI]

    "Here are today's top AI developments: 1. OpenAI released... 2. Google announced... [summary with 5 sources]"

    💰 Cost: ~32 credits (model) + 20 (Exa) = 52 credits

    🤖 Model Capabilities

    AGENTIC
    Models with Native Function Calling:

    OpenAI:

    GPT-5, GPT-5.2, GPT-5 Nano, GPT-OSS-120B

    Anthropic:

    Claude 4.5 Sonnet, Claude Opus 4.5

    Google:

    Gemini 2.0 Flash, Gemini 3 Pro Preview

    Others:

    Azure OpenAI models (EU routing)

    ✅ Can use native search_web() and fetch_url() autonomously

    Models That Need Exa Plugin:

    DeepSeek:

    DeepSeek R1, DeepSeek Chat

    Qwen:

    Qwen QwQ, Qwen2.5, Qwen3

    Meta:

    Llama 3.3, Llama 3.1 models

    Others:

    Mistral, Mixtral, older models

    ✅ Enable Web Search toggle to use Exa plugin

    ⚠️ Why Results May Differ

    The same question with the same model can produce different results depending on which search system is active:

    Native Agentic (Toggle OFF):

    • • Uses Brave Search (traditional web index)
    • • Gets 10 results (more comprehensive)
    • • Can read full pages for deep context
    • • May search multiple times with refined queries
    • • More thorough but may take longer

    Exa Plugin (Toggle ON):

    • • Uses Exa AI (semantic, AI-optimized)
    • • Gets 5 results (focused)
    • • Only sees snippets (no full pages)
    • • Single search per question
    • • Faster, more predictable

    💡 Recommendation:

    For AGENTIC models, we recommend keeping the toggle OFF to unleash their full research capabilities. For non-AGENTIC models, turn the toggle ON to enable Exa search.

    🚀 Getting Started

    For AGENTIC Models (GPT-5, Claude 4.5+, Gemini 3+):

    1. 1. Select an AGENTIC model (look for the purple badge)
    2. 2. Keep Web Search toggle OFF (default)
    3. 3. Ask questions that need research
    4. 4. Watch the model autonomously search and fetch URLs

    For Non-AGENTIC Models (DeepSeek, Qwen, Llama):

    1. 1. Go to Settings → My Custom Models
    2. 2. Find your model and toggle Web Search ON
    3. 3. Ask questions about current events
    4. 4. Model will use Exa plugin for search

    🎯 Pro Tip:

    Try the SAME question with toggle OFF (native) and toggle ON (Exa) to see which gives better results for your use case. Deep research usually benefits from native autonomy, while quick fact-checks work well with Exa.

    📖 Learn More

    AGENTIC Capabilities Guide

    Learn about all native tools: memory, knowledge, chat history, and web research

    Related Articles

    Basics

    AGENTIC AI: Models That Think, Remember & Research Autonomously

    Discover how frontier AI models can now manage your memories, search knowledge bases, and access chat history—without you asking.

    Read more
    Basics

    NEW: Google Drive & OneDrive Integration — Access Your Files Directly in Chat

    Attach documents from your cloud storage with one click. Connect Google Drive and Microsoft OneDrive to AI·Collab for seamless file access.

    Read more
    Basics

    Multiple Models in Tabs: Compare, Learn & Merge Results

    Use multiple AI models simultaneously in separate tabs. Compare different perspectives, verify information, and merge the best insights into one comprehensive answer.

    Read more

    Ready to Experience 300+ AI Models?

    Get started today. Access models from OpenAI, Google, Anthropic, Grok and more.

    GDPR compliant · Zero data retention · Cancel anytime