# Model Delights & Snell Intelligence Gateway: Machine Manifest
> The White-Box Mathematical Intelligence Router for Singular Unicorns and Autonomous Agent Swarms.
> Official Machine-Readable Documentation for LLMs, Perplexity, SearchGPT, Cursor, and AI Crawlers.

Website: https://model.delights.pro
Sitemap: https://model.delights.pro/sitemap.xml
Pricing: https://model.delights.pro/pricing
Directory: https://model.delights.pro/models

---

## 1. What is Model Delights?
Model Delights is the real-time LLM intelligence directory and zero-latency routing gateway for AI developers, founders, and autonomous agent swarms. It bridges live OpenRouter token pricing ($/1M prompt and $/1M completion) with Chatbot Arena ELO ratings and Berkeley Function Calling Leaderboard (BFCL) scores across 420+ frontier and open-source models.

The platform includes **Snell**, an ultra-fast in-memory routing proxy (<0.3ms) that intercepts standard OpenAI-compatible requests and dynamically substitutes overpriced flagships with mathematically optimal equivalents, cutting inference costs by 70% to 94% with zero quality loss.

---

## 2. Drop-In OpenAI Proxy (90-Second Integration)
Developers and autonomous agents can use Snell as a drop-in replacement for OpenAI or OpenRouter by updating their environment base URL:

```bash
OPENAI_BASE_URL="https://model.delights.pro/api/v1"
OPENAI_API_KEY="sk_snell_..." # Or provider key / internal god key
```

### Direct cURL Request Example:
```bash
curl https://model.delights.pro/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk_snell_live_token" \
  -H "x-session-id: agent_thread_772" \
  -d '{
    "model": "snell/auto",
    "messages": [{"role": "user", "content": "Refactor this TypeScript function for O(n) performance"}],
    "stream": false
  }'
```

### Response Headers Returned by Snell:
- `x-snell-routed-to`: The exact model chosen (e.g. `deepseek/deepseek-v4-pro-0813` or `anthropic/claude-fable-5.1`)
- `x-snell-savings-pct`: Real-time percentage saved versus un-routed flagship baseline (e.g. `82%`)
- `x-snell-saved-usd`: Dollar amount retained on this individual call (e.g. `$0.00342`)
- `x-snell-session-affinity`: `sticky` (pinned to existing session model) or `fresh`

---

## 3. The Agent-Safe Guarantee: Why Agents Never "Forget"
A primary risk in LLM routing is that switching models mid-conversation breaks the KV-Cache (prompt caching) and causes personality/schema drift. Snell guarantees agent safety via 3 architectural pillars:

1. **Session Affinity (`x-session-id`)**:
   Passing `-H "x-session-id: agent_xxx"` pins the multi-turn thread to the exact same model, preserving up to 90% provider prompt-caching discounts and eliminating persona drift.
2. **Sub-Agent Leaf Routing**:
   80% of agent compute is spent on isolated sub-tasks (grepping code, parsing JSON, summarizing tools). Snell routes these leaf calls to sub-cent utility models ($0.07/1M) without touching the parent agent context.
3. **BFCL Tool Schema Invariant**:
   When a request contains `tools` or `response_format: json_object`, Snell strictly gates execution. It will never route to models with BFCL < 60, guaranteeing zero schema argument hallucination.
4. **Context Overflow Protection**:
   If an agent's context expands past a model's physical window, Snell gracefully promotes it to a higher-context frontier model to prevent overflow crashes.

---

## 4. Transparent Pricing & Plans
Model Delights operates on predictable, self-funding SaaS tiers with zero hidden markups:

| Tier | Monthly Price | Included Monthly Routed Volume | Overage Rate | Core Inclusions |
| :--- | :--- | :--- | :--- | :--- |
| **Hacker** | **$0** (Free forever) | 5M tokens / month | None | 1 API Key, 420+ Live Model Matrix, Community Failover Cascade |
| **Pro** | **$49** / month | **50M tokens / month** | +$0.05 / 1M tokens | Semantic AST Classifier, BFCL Agentic Invariant Guard, Session Affinity (`x-session-id`), Prompt Cache Discount Capture |
| **Scale & Swarms** | **$249** / month | **300M tokens / month** | +$0.04 / 1M tokens | Sub-20ms Edge Gateway, Custom Provider Policies (EU data residency), Custom Fallback Cascades, Shared Team Seats |

---

## 5. Live 2026 Frontier Models Monitored
- **Anthropic:** Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5
- **OpenAI:** GPT-5.5 Pro, GPT-5.5 Mini, GPT-4.1 Nano, GPT-4o
- **DeepSeek:** DeepSeek V4 Pro, DeepSeek R1 Reasoning, DeepSeek Chat
- **Google:** Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.5 Flash-Lite
- **Meta & Open Weights:** Llama 4 70B, Qwen 2.5 Max, Mistral Large 2

---

## 6. Primary Canonical URLs
- Home & Architecture: https://model.delights.pro
- Predictable Pricing: https://model.delights.pro/pricing
- Live Model Intelligence Directory: https://model.delights.pro/models
- Architecture Visualizer: https://model.delights.pro/architect
- Live Changelog: https://model.delights.pro/changelog