OmniRoute/.omo/deepseek-web-integration/COMPARISON_MATRIX.md

COMPARISON_MATRIX.md - DeepSeek vs Claude vs ChatGPT Web APIs

Comparison Overview

Dimension	DeepSeek	Claude.ai	ChatGPT
Base URL	api.deepseek.com	claude.ai	chat.openai.com
Streaming	SSE	SSE	SSE
Auth Method	Cookie-based	Cookie-based	Cookie-based
Session TTL	24h-30d	~7d	~24h
Rate Limit	60 req/min, 100K tokens/day	40 conv/day	Unknown (strict)
Concurrent Limit	10-50 req	1-2 concurrent	1 concurrent
Error Handling	JSON errors + SSE errors	JSON errors	JSON errors
Model Selection	Parameter: model	Auto-selected	Auto-selected
Conversation Model	UUID per conversation	UUID per conversation	UUID per conversation

---

API Endpoint Comparison

DeepSeek

POST /api/v0/chat/completions
Headers: Cookie, Content-Type
Body: {"model": "deepseek-v4-flash", "messages": [...], "stream": true}

Claude.ai

POST /api/organizations/{org_id}/chat_conversations/{conv_id}/completion
Headers: Cookie, anthropic-device-id, anthropic-client-platform: web_claude_ai
Body: {"prompt": "...", "attachments": [...], "organization_id": "..."}

ChatGPT

POST /backend-api/conversation
Headers: Cookie, authorization
Body: {"action": "next", "messages": [...], "model": "text-davinci-004-code"}

---

Authentication Mechanisms

DeepSeek

• Method: Browser cookies (_deepseek_session, __Secure-deepseek-id)
• Persistence: File-based or in-memory cookie jar
• Refresh: Automatic via activity
• Expiry: 24-30 days inactivity
• Challenge: Sessions may rotate or refresh unpredictably

Claude.ai

• Method: Browser cookies (sessionKey) + Device ID (UUID)
• Persistence: File-based or in-memory
• Refresh: Requires periodictouches (requests)
• Expiry: ~7 days absolute
• Challenge: Cloudflare cf_clearance cookie required

ChatGPT

• Method: Browser cookies + Bearer token in header
• Persistence: File-based
• Refresh: Via /auth/session endpoint
• Expiry: Varies (1-30 days)
• Challenge: Token rotation, Cloudflare protection, strictest rate limiting

---

Streaming Format Comparison

DeepSeek

data: {"choices":[{"delta":{"content":"Hello"}}],"model":"deepseek-v4"}
data: {"choices":[{"delta":{"content":" world"}}],"model":"deepseek-v4"}
data: [DONE]

• Protocol: SSE (text/event-stream)
• Format: data: {JSON}
• End Marker: data: [DONE]

Claude.ai

event: message_delta
data: {"type":"content_block_delta","delta":{"type":"text_delta","text":"Hello"}}

event: message_stop
data: {"type":"message_delta_stop"}

• Protocol: SSE with named events
• Format: event: {name} + data: {JSON}
• End Marker: event: message_stop

ChatGPT

data: {"message":{"content":[{"content_type":"text","parts":["Hello"]}]}}
data: [DONE]

• Protocol: SSE
• Format: data: {JSON} (full message state each time)
• End Marker: data: [DONE]

---

Error Handling Patterns

DeepSeek

HTTP Errors:

• 400: Invalid request
• 401: Unauthorized
• 429: Rate limited
• 500: Server error
• 503: Service unavailable

SSE Errors: JSON error objects within stream

Recovery: Exponential backoff, retry with limits

Claude.ai

HTTP Errors:

• 400: Invalid request
• 401: Session expired
• 429: Rate limited
• 500: Server error

SSE Errors: Error events (e.g., event: error)

Recovery: Longer backoff times (Claude is stricter)

ChatGPT

HTTP Errors:

• 401: Unauthorized
• 429: Rate limited (very strict)
• 500: Server error

SSE Errors: JSON objects with error field

Recovery: Very long backoffs required (1min+)

---

Session Management Comparison

DeepSeek

• Multi-Tab: Shared session
• Concurrent Requests: 10-50 allowed
• Conversation Limit: Many per session
• Session Refresh: Automatic on activity
• Logout: Explicit endpoint or cookie delete

Claude.ai

• Multi-Tab: Shared session
• Concurrent Requests: 1-2 allowed (strict)
• Conversation Limit: ~40 per day (usage-based)
• Session Refresh: Periodic touches required
• Logout: Via API endpoint

ChatGPT

• Multi-Tab: Shared session
• Concurrent Requests: 1 only (strictest)
• Conversation Limit: Unlimited per day (rate limited)
• Session Refresh: Via /auth/session endpoint
• Logout: Via logout endpoint

---

Message & Conversation Format

DeepSeek

{
  "role": "user|assistant|system",
  "content": "Text content"
}

• Simple text messages
• No attachment support
• No image support (in web wrapper)
• System prompt as role: "system"

Claude.ai

{
  "type": "text",
  "text": "Message content",
  "attachments": [
    {"id": "file-123", "name": "document.pdf"}
  ]
}

• Complex objects
• Attachment support
• Image/file support
• Organization ID required

ChatGPT

{
  "id": "msg-123",
  "author": {"role": "user|assistant"},
  "content": [
    {"content_type": "text", "parts": ["Hello"]}
  ]
}

• Nested content blocks
• Multiple content types
• Complex metadata
• Model parameter required

---

Rate Limiting Comparison

DeepSeek

• Requests/Min: 60
• Tokens/Day: 100,000-1M (tier-dependent)
• Concurrent: 10-50
• Headers: X-RateLimit-Limit-Requests, X-RateLimit-Remaining-Requests, Retry-After
• Behavior: 429 with Retry-After

Claude.ai

• Requests/Min: ~40
• Conversations/Day: ~40
• Concurrent: 1-2
• Headers: Not standard
• Behavior: 429 with very long backoff

ChatGPT

• Requests/Min: Unknown (very strict)
• Daily Limit: Message count + model tier
• Concurrent: 1 only
• Headers: Not standard
• Behavior: 429 with long backoff (1min+)

---

Model & Parameter Comparison

DeepSeek

Models: deepseek-v4-flash, deepseek-v4-pro, deepseek-r1, deepseek-v3 Parameters:

• model (required)
• messages (required)
• stream (optional, default: false)
• temperature (0-2, default: 1)
• max_tokens (optional)
• reasoning_effort (low, medium, high)
• top_p (0-1, default: 1)

Claude.ai

Models: Auto-selected by Claude.ai (no parameter) Parameters:

• prompt (required)
• model (hidden, auto-selected)
• attachments (optional)
• temperature (0-1, default: 1)
• system (system prompt, optional)

ChatGPT

Models: text-davinci-004-code (hidden from web UI) Parameters:

• model (hidden, auto-selected)
• messages (required)
• temperature (0-2, default: 1)
• max_tokens (optional)
• top_p (0-1, default: 1)

---

Implementation Difficulty Ranking

Easiest to Hardest

1. DeepSeek ⭐⭐ (Easiest)

   • Clear API structure
   • Standard SSE format
   • Reasonable rate limits
   • Good concurrency support

2. Claude.ai ⭐⭐⭐ (Medium)

   • Strict concurrency (1-2)
   • Cloudflare protection
   • Complex attachment handling
   • Session rotation

3. ChatGPT ⭐⭐⭐⭐⭐ (Hardest)

   • Strictest rate limiting (1 concurrent)
   • Token rotation required
   • No official API exposed
   • Cloudflare + additional protections
   • Very long backoffs needed

---

Unique Challenges by Provider

DeepSeek

• Session cookie format changes
• Reasoning effort parameter (new)
• Token usage tracking

Claude.ai

• Cloudflare cf_clearance cookie required
• Device ID must persist
• Conversation limit enforcement
• Attachment upload handling

ChatGPT

• Strictest concurrency (1 only)
• Longest rate limit backoffs
• Token expiration & refresh
• Most aggressive bot detection
• No streaming response for initial request

---

Recommended Web Wrapper Approach

For DeepSeek

1. Use cookie jar (tough-cookie)
2. Parse SSE stream line-by-line
3. Implement backoff for 429/500
4. Queue concurrent requests (limit to 5-10)
5. Refresh session every 24h

For Claude.ai

1. Use cookie jar + device ID persistence
2. Handle Cloudflare challenge
3. Limit to 1-2 concurrent requests
4. Parse named SSE events
5. Handle attachment uploads

For ChatGPT

1. Strict 1 concurrent request limit
2. Implement 1-5min backoff for 429
3. Parse SSE with full message state
4. Refresh token regularly
5. Expect bot detection responses

---

Shared Patterns Across All Three

✅ All use SSE for streaming
✅ All use cookie-based authentication
✅ All have session TTL (1-30 days)
✅ All support messages array format
✅ All have rate limiting
✅ All require User-Agent header
✅ All use 401 for auth failure

❌ All have different concurrent limits
❌ All have different rate limits
❌ All have different streaming formats
❌ All have different error recovery strategies