# Deep Research via OpenRouter API - Comprehensive Guide
This report documents all AI models supporting Deep Research capabilities through the OpenRouter API, including parameters needed to trigger proper deep research with web search.
---
## Quick Reference: Models Supporting Deep Research
### Dropdown List of Deep Research Models
```javascript
const DEEP_RESEARCH_MODELS = [
// Anthropic Claude 4.5 Family
{ id: "anthropic/claude-opus-4.5", name: "Claude Opus 4.5", features: ["extended_thinking", "web_search"] },
{ id: "anthropic/claude-sonnet-4.5", name: "Claude Sonnet 4.5", features: ["extended_thinking", "web_search"] },
{ id: "anthropic/claude-haiku-4.5", name: "Claude Haiku 4.5", features: ["extended_thinking", "web_search"] },
// Perplexity Sonar
{ id: "perplexity/sonar-deep-research", name: "Perplexity Sonar Deep Research", features: ["multi_step_retrieval", "auto_citations"] },
{ id: "perplexity/sonar-reasoning-pro", name: "Perplexity Sonar Reasoning Pro", features: ["reasoning", "citations"] },
// OpenAI / ChatGPT
{ id: "openai/chatgpt-4o-latest", name: "ChatGPT-4o", features: ["reasoning", "web_search"] },
{ id: "openai/gpt-5.2", name: "GPT-5.2", features: ["reasoning", "web_search"] },
{ id: "openai/o3", name: "OpenAI o3", features: ["deep_reasoning", "web_search"] },
// Google Gemini
{ id: "google/gemini-3-pro-preview", name: "Gemini 3 Pro Preview", features: ["thinking", "grounding"] },
{ id: "google/gemini-3-flash-preview", name: "Gemini 3 Flash Preview", features: ["thinking", "grounding"] },
// xAI Grok
{ id: "x-ai/grok-4.1", name: "Grok 4.1", features: ["reasoning", "web_search", "x_search"] },
{ id: "x-ai/grok-4.1-fast", name: "Grok 4.1 Fast", features: ["reasoning", "web_search", "x_search"] },
{ id: "x-ai/grok-4.1-fast:free", name: "Grok 4.1 Fast (Free)", features: ["reasoning", "web_search"] },
];
```
---
## 1. Perplexity Sonar Deep Research
### Model ID
```
perplexity/sonar-deep-research
```
### Overview
- **Context Window**: 128,000 tokens
- **Pricing**: $2/M input, $8/M output, $5/K web search
- **Specialty**: Multi-step retrieval, synthesis, and reasoning across complex topics
### API Parameters
| Parameter | Type | Values | Description |
|-----------|------|--------|-------------|
| `model` | string | `perplexity/sonar-deep-research` | Model identifier |
| `return_citations` | boolean | `true` | Return source citations |
| `return_related_questions` | boolean | `true` | Return related queries |
| `search_recency_filter` | string | `day`, `week`, `month`, `year` | Filter by date |
### Request Example
```javascript
{
"model": "perplexity/sonar-deep-research",
"messages": [
{
"role": "user",
"content": "Conduct comprehensive research on quantum computing applications in finance"
}
],
"return_citations": true,
"return_related_questions": true,
"search_recency_filter": "month"
}
```
### Key Features
- **Autonomous multi-step retrieval**: Automatically searches, reads, and evaluates sources
- **Report generation**: Synthesizes findings into comprehensive reports
- **Citation support**: Returns sources with titles and URLs
- **Domain specialization**: Finance, technology, health, current events
---
## 2. ChatGPT-4o / OpenAI Models
### Model IDs
```
openai/chatgpt-4o-latest
openai/gpt-5.2
openai/o3
```
### Overview
- **Context Window**: 128,000 tokens
- **Pricing**: $0.005/M input, $0.015/M output
- **Web Search**: Via OpenRouter plugins or native tool
### API Parameters
| Parameter | Type | Values | Description |
|-----------|------|--------|-------------|
| `model` | string | `openai/chatgpt-4o-latest` | Model identifier |
| `reasoning` | object | `{"effort": "high"}` | Reasoning effort level |
| `reasoning.effort` | string | `low`, `medium`, `high`, `xhigh` | Thinking depth |
| `plugins` | array | `[{"id": "web"}]` | Enable web search |
| `plugins[].max_results` | integer | `1-20` (default: 5) | Search result count |
### Request Example - Standard
```javascript
{
"model": "openai/chatgpt-4o-latest",
"messages": [
{
"role": "user",
"content": "Research the latest developments in AI regulation"
}
],
"reasoning": {
"effort": "high"
},
"plugins": [
{
"id": "web",
"max_results": 10
}
]
}
```
### Request Example - GPT-5.2 with Extended Reasoning
```javascript
{
"model": "openai/gpt-5.2",
"messages": [
{"role": "user", "content": "Deep analysis of semiconductor supply chain"}
],
"reasoning": {
"effort": "xhigh" // Maximum reasoning depth
},
"plugins": [
{
"id": "web",
"max_results": 15
}
]
}
```
### Key Limitations
> **Note**: OpenAI's Deep Research feature (available in ChatGPT Pro) is NOT available via API. The API supports reasoning + web search, but not the full Deep Research agent.
---
## 3. Google Gemini 3 Pro
### Model IDs
```
google/gemini-3-pro-preview
google/gemini-3-flash-preview
```
### Overview
- **Context Window**: 1M tokens
- **Pricing**: $0.75/M input, $3/M output
- **Specialty**: Multimodal with thinking capabilities
### API Parameters
| Parameter | Type | Values | Description |
|-----------|------|--------|-------------|
| `model` | string | `google/gemini-3-pro-preview` | Model identifier |
| `reasoning` | object | `{"effort": "high"}` | Enable thinking |
| `reasoning.effort` | string | `minimal`, `low`, `medium`, `high` | Thinking level |
| `plugins` | array | `[{"id": "web"}]` | Web grounding |
### Request Example
```javascript
{
"model": "google/gemini-3-pro-preview",
"messages": [
{
"role": "user",
"content": "Research and analyze renewable energy trends"
}
],
"reasoning": {
"effort": "high"
},
"plugins": [
{
"id": "web",
"max_results": 10
}
]
}
```
### Important Notes
- **Reasoning Details Required**: Gemini models require `reasoning_details` to be preserved in multi-turn conversations
- **Thought Signature**: Function calls require `thought_signature` in functionCall parts
---
## 4. Anthropic Claude 4.5 Family
### Model IDs
```
anthropic/claude-opus-4.5 # Frontier model ($3/M input, $15/M output)
anthropic/claude-sonnet-4.5 # Balanced ($1/M input, $3/M output)
anthropic/claude-haiku-4.5 # Fast ($0.25/M input, $0.75/M output)
```
### Overview
- **Context Window**: 200,000 tokens
- **Extended Thinking**: Up to 128K tokens budget
- **Web Search**: Native `web_search_20250305` tool
### API Parameters
| Parameter | Type | Values | Description |
|-----------|------|--------|-------------|
| `model` | string | `anthropic/claude-opus-4.5` | Model identifier |
| `reasoning` | object | `{"effort": "high"}` | Extended thinking control |
| `reasoning.effort` | string | `low`, `medium`, `high` | Thinking depth |
| `reasoning.max_tokens` | integer | `1024-128000` | Thinking budget in tokens |
| `verbosity` | string | `low`, `medium`, `high` | Token efficiency control |
| `plugins` | array | `[{"id": "web"}]` | Enable native search |
### Request Example - Extended Thinking + Web Search
```javascript
{
"model": "anthropic/claude-opus-4.5",
"messages": [
{
"role": "user",
"content": "Conduct deep research on AI safety implications"
}
],
"reasoning": {
"effort": "high",
"max_tokens": 32000 // Thinking budget
},
"plugins": [
{
"id": "web",
"max_results": 10
}
]
}
```
### Native Web Search Tool (Direct Anthropic Format)
For direct Anthropic API calls (not OpenRouter):
```javascript
{
"model": "claude-opus-4-5-20251101",
"max_tokens": 64000,
"thinking": {
"type": "enabled",
"budget_tokens": 32000
},
"tools": [{
"type": "web_search_20250305",
"name": "web_search",
"max_uses": 10
}],
"messages": [{"role": "user", "content": "Research topic"}]
}
```
### Key Features
- **Extended Thinking**: Deep multi-step reasoning visible in response
- **Verbosity Control**: `low`, `medium`, `high` for token efficiency
- **SWE-bench**: Opus 4.5 achieves 74.5% on SWE-bench Verified
---
## 5. xAI Grok 4.1
### Model IDs
```
x-ai/grok-4.1
x-ai/grok-4.1-fast
x-ai/grok-4.1-fast:free # Free tier available
x-ai/grok-4.1-fast:online # Shorthand for web search
```
### Overview
- **Context Window**: 2M tokens
- **Pricing**: $0.50/M input, $1.50/M output, $4/K web searches
- **Specialty**: X platform search + web search
### API Parameters
| Parameter | Type | Values | Description |
|-----------|------|--------|-------------|
| `model` | string | `x-ai/grok-4.1-fast` | Model identifier |
| `reasoning` | object | `{"enabled": true}` | Enable reasoning |
| `plugins` | array | Web/X search config | Search plugins |
| `plugins[].id` | string | `web`, `x_search` | Plugin type |
| `plugins[].max_results` | integer | `1-20` (default: 5) | Result count |
| `plugins[].engine` | string | `exa`, `native` | Search engine |
| `web_search_options.context_size` | string | `low`, `medium`, `high` | Context depth |
### Request Example - Full Deep Research
```javascript
{
"model": "x-ai/grok-4.1-fast",
"messages": [
{
"role": "user",
"content": "Research the competitive landscape of electric vehicle batteries"
}
],
"reasoning": {
"enabled": true
},
"plugins": [
{
"id": "web",
"max_results": 15,
"engine": "exa"
}
],
"web_search_options": {
"context_size": "high" // Extensive context for deep research
}
}
```
### Shorthand Method `:online`
```javascript
{
"model": "x-ai/grok-4.1-fast:online", // :online enables web search
"messages": [{"role": "user", "content": "Research topic"}],
"reasoning": {"enabled": true}
}
```
### X Platform Search (x_search)
```javascript
{
"model": "x-ai/grok-4.1-fast",
"messages": [{"role": "user", "content": "Summarize Elon Musk's recent AI views"}],
"plugins": [
{
"id": "x_search",
"allowed_x_handles": ["elonmusk"],
"from_date": "2025-12-01",
"to_date": "2026-02-02"
}
]
}
```
---
## OpenRouter Universal Parameters
### Web Search Plugin (Works Across All Models)
```javascript
"plugins": [
{
"id": "web",
"max_results": 10, // 1-20 (default: 5)
"engine": "exa", // "exa" or "native" (for Anthropic/OpenAI)
"search_prompt": "Custom instruction for incorporating results"
}
]
```
### Reasoning Parameters (Unified)
OpenRouter normalizes reasoning parameters across providers:
```javascript
// Effort-based (works across providers)
"reasoning": {
"effort": "high" // low, medium, high, xhigh
}
// Token budget (for Claude/extended thinking)
"reasoning": {
"max_tokens": 32000
}
```
### Response Format with Citations
```javascript
{
"message": {
"content": "Research findings...",
"annotations": [
{
"type": "url_citation",
"url_citation": {
"url": "https://source.com/article",
"title": "Article Title",
"content": "Snippet from source",
"start_index": 100,
"end_index": 200
}
}
]
}
}
```
---
## Comparison Matrix
| Feature | Perplexity Sonar | ChatGPT-4o | Gemini 3 Pro | Claude 4.5 | Grok 4.1 |
|---------|------------------|------------|--------------|------------|----------|
| **Model ID** | `perplexity/sonar-deep-research` | `openai/chatgpt-4o-latest` | `google/gemini-3-pro-preview` | `anthropic/claude-opus-4.5` | `x-ai/grok-4.1-fast` |
| **Extended Thinking** | ✅ Built-in | ✅ `reasoning.effort` | ✅ `reasoning.effort` | ✅ 128K budget | ✅ `reasoning.enabled` |
| **Web Search** | ✅ Auto | ✅ Plugin | ✅ Plugin | ✅ Native/Plugin | ✅ Plugin + X Search |
| **Citations** | ✅ Auto | ✅ Annotations | ❓ Limited | ✅ Tool results | ✅ Annotations |
| **Context** | 128K | 128K | 1M | 200K | 2M |
| **Best For** | Research reports | General research | Multimodal | Deep analysis | Real-time X data |
| **Price (input/output)** | $2/$8/M | $0.005/$0.015/M | $0.75/$3/M | $3/$15/M (Opus) | $0.50/$1.50/M |
---
## Implementation Recommendations
### 1. For Comprehensive Research Reports
```javascript
// Perplexity Sonar Deep Research
"model": "perplexity/sonar-deep-research",
"return_citations": true
```
### 2. For Cost-Effective Research
```javascript
// Grok 4.1 Fast with web search
"model": "x-ai/grok-4.1-fast",
"plugins": [{"id": "web", "max_results": 10}],
"reasoning": {"enabled": true}
```
### 3. For Maximum Reasoning Depth
```javascript
// Claude Opus 4.5 with extended thinking
"model": "anthropic/claude-opus-4.5",
"reasoning": {"effort": "high", "max_tokens": 64000},
"plugins": [{"id": "web"}]
```
### 4. For Real-Time Social Data
```javascript
// Grok with X search
"model": "x-ai/grok-4.1-fast",
"plugins": [{"id": "x_search", "from_date": "2026-01-01"}]
```
### 5. For Multimodal Research
```javascript
// Gemini 3 Pro
"model": "google/gemini-3-pro-preview",
"reasoning": {"effort": "high"},
"plugins": [{"id": "web"}]
```
---
## Extending Web Search Context for Deep Research
This section documents all mechanisms to increase the amount of web search context available to AI models.
### OpenRouter Web Plugin Parameters
#### `max_results` (1-20)
Controls the number of search results returned. Default is 5.
```javascript
"plugins": [
{
"id": "web",
"max_results": 20 // Maximum: 20 results per search
}
]
```
**Pricing Impact:**
- Exa Search: $4 per 1,000 results
- With `max_results: 20` = $0.08 per request
#### `search_prompt` (Custom Instructions)
Add custom instructions for how the model should incorporate results:
```javascript
"plugins": [
{
"id": "web",
"max_results": 15,
"search_prompt": "Analyze each source critically. Incorporate these results with proper citations, cross-reference facts, and identify consensus vs. outliers:"
}
]
```
---
### Exa Search API Parameters (Deep Mode)
When using Exa via OpenRouter or directly, use these parameters for maximum context:
#### Search Types
| Type | Description | Use Case |
|------|-------------|----------|
| `fast` | Quick results, lower latency | Basic queries |
| `auto` | Balanced (default) | Most use cases |
| `deep` | Maximum depth, higher quality | Deep research |
#### Content Retrieval Options
```javascript
// Maximum context configuration
{
"type": "deep",
"num_results": 20, // Up to 20 results
"text": {
"max_characters": 20000 // Up to 20K chars per result
},
"additional_queries": [ // Query variations for broader coverage
"alternative phrasing 1",
"alternative phrasing 2"
],
"livecrawl": "always" // Force fresh content
}
```
#### Context String Mode (For RAG)
```javascript
// Single string with aggregated context
{
"type": "deep",
"num_results": 20,
"context": {
"max_characters": 50000 // Up to 50K chars total
}
}
```
#### Verbosity Levels
```javascript
"text": {
"verbosity": "full", // "compact", "standard", "full"
"includeSections": ["body", "header"],
"excludeSections": ["navigation", "footer", "sidebar"]
}
```
---
### Anthropic Claude `max_uses` Parameter
Controls the number of web searches Claude can perform:
```javascript
// Direct Anthropic API
{
"tools": [{
"type": "web_search_20250305",
"name": "web_search",
"max_uses": 20 // Default: 5, maximum varies
}]
}
```
**Key Insight**: Claude can perform *iterative searches*, refining queries based on previous results. Higher `max_uses` enables deeper multi-step research.
---
### xAI Grok `search_context_size` Parameter
Controls the amount of context extracted from search results:
```javascript
{
"model": "x-ai/grok-4.1-fast",
"web_search_options": {
"search_context_size": "high" // "low", "medium" (default), "high"
}
}
```
| Level | Description | Token Impact |
|-------|-------------|--------------|
| `low` | Minimal context, basic queries | Low |
| `medium` | Moderate context (default) | Medium |
| `high` | Extensive context, **ideal for deep research** | High |
---
### Perplexity Sonar Extended Parameters
Perplexity handles context automatically but provides these controls:
```javascript
{
"model": "perplexity/sonar-deep-research",
"return_citations": true,
"return_related_questions": true,
"search_recency_filter": "week", // Narrow time window for fresher content
"search_domain_filter": [ // Focus on specific domains
"arxiv.org",
"github.com",
"nature.com"
]
}
```
**Note**: Perplexity API currently returns 3-5 citations by default. Full citation access requires elevated tier access.
---
### Maximum Context Configuration Examples
#### Example 1: OpenRouter + Exa Maximum Context
```javascript
{
"model": "anthropic/claude-opus-4.5",
"messages": [{"role": "user", "content": "Deep research on topic"}],
"reasoning": {
"effort": "high",
"max_tokens": 64000
},
"plugins": [
{
"id": "web",
"max_results": 20, // Maximum results
"engine": "exa",
"search_prompt": "Provide comprehensive analysis with extensive citations:"
}
]
}
```
#### Example 2: xAI Grok Maximum Context
```javascript
{
"model": "x-ai/grok-4.1-fast",
"messages": [{"role": "user", "content": "Research topic"}],
"reasoning": {"enabled": true},
"plugins": [
{
"id": "web",
"max_results": 20,
"engine": "exa"
}
],
"web_search_options": {
"search_context_size": "high"
}
}
```
#### Example 3: Claude Native Maximum Search
```javascript
// Direct Anthropic API (not OpenRouter)
{
"model": "claude-opus-4-5-20251101",
"max_tokens": 128000,
"thinking": {
"type": "enabled",
"budget_tokens": 64000
},
"tools": [{
"type": "web_search_20250305",
"name": "web_search",
"max_uses": 20
}],
"messages": [{"role": "user", "content": "Comprehensive research"}]
}
```
---
### Context Expansion Summary Table
| Provider | Parameter | Default | Maximum | Notes |
|----------|-----------|---------|---------|-------|
| OpenRouter (Exa) | `max_results` | 5 | 20 | $0.004/result |
| Exa Direct | `num_results` | 10 | 20+ | Higher with enterprise |
| Exa Direct | `text.max_characters` | 15000 | 20000+ | Per result |
| Exa Direct | `context.max_characters` | - | 50000+ | Total context |
| Anthropic | `max_uses` | 5 | varies | Iterative searches |
| Grok | `search_context_size` | medium | high | Extensive context |
| Perplexity | citations | 3-5 | varies | Elevated access needed |
---
## Best Practices
1. **Set `max_results: 10-15`** for comprehensive research
2. **Enable reasoning** for better analysis and synthesis
3. **Use `context_size: "high"`** (Grok) for extensive context
4. **Preserve `reasoning_details`** in multi-turn conversations (Gemini)
5. **Stream responses** for long research tasks: `"stream": true`
6. **Set timeout to `None`** for extended thinking operations (32K+ tokens)
---
## Pricing Summary
| Model | Input | Output | Web Search |
|-------|-------|--------|------------|
| Perplexity Sonar | $2/M | $8/M | $5/K searches |
| ChatGPT-4o | $0.005/M | $0.015/M | Included in plugins |
| Gemini 3 Pro | $0.75/M | $3/M | Plugin cost |
| Claude Opus 4.5 | $3/M | $15/M | Plugin cost |
| Claude Sonnet 4.5 | $1/M | $3/M | Plugin cost |
| Grok 4.1 Fast | $0.50/M | $1.50/M | $4/K results (Exa) |
---
## Version History
| Version | Date | Changes |
|---------|------|---------|
| 1.0.0 | 2026-02-02 | Initial comprehensive report |
## References
- [OpenRouter Web Search Docs](https://openrouter.ai/docs/features/web-search)
- [Perplexity Sonar Deep Research](https://openrouter.ai/perplexity/sonar-deep-research)
- [Anthropic Extended Thinking](https://docs.claude.com/en/docs/build-with-claude/extended-thinking)
- [xAI Grok API](https://docs.x.ai/docs)
- [GrokResearch.md](./GrokResearch.md) - Local reference for Grok parameters
