How We Built Multi-AI Model Routing (And Why It Matters)
Most AI builders use one model. We use five. Here's how our intelligent routing system picks the right AI for every task.
Every AI website builder picks a model and sticks with it. Lovable uses Claude. Bolt uses a mix. v0 uses their own.
We use **five models simultaneously** — and built a routing system that picks the right one for every task. Here's why, and how it works.
The Problem with Single-Model AI
When you depend on one AI provider, three things can go wrong:
1. **Outages** — Claude goes down, your product goes down
2. **Rate limits** — hit the API ceiling during peak hours, users wait
3. **Capability gaps** — no single model is best at everything
We hit all three. So we built something better.
The Architecture
Our system has three layers:
Layer 1: Task Classification
When you send a message, a lightweight classifier (Haiku) categorizes it in under 200ms:
- **question** → simple answer, no code changes
- **style_change** → CSS/color/font tweaks
- **bug_fix** → something broke, fix it
- **component_create** → build a new section or component
- **feature** → add functionality (forms, navigation, etc.)
- **refactor** → restructure existing code
- **architecture** → major structural changes
For unambiguous prompts, we skip the classifier entirely and use heuristics — pattern matching on keywords and message length. This saves a round-trip for ~60% of messages.
Layer 2: Model Selection
Based on the task complexity:
| Task Type | Model | Why |
|---|---|---|
| Questions, simple edits | **Haiku** | Fast (sub-second), cheap, good enough |
| Standard pages, components | **Sonnet** | Best balance of speed and quality |
| Complex architecture, multi-file | **Opus** | Deepest reasoning, worth the cost |
Pro plan users get access to all three tiers. Free and Starter users get Haiku + Sonnet.
Layer 3: Fallback Chains
If the primary model fails (timeout, rate limit, error), we automatically try the next provider:
Claude Sonnet → DeepSeek V3 → Gemini Flash → GPT-4o Mini
The user never sees an error. The fallback adds 1-2 seconds of latency, but the generation completes successfully. In practice, fallbacks trigger about 3% of the time — rare enough not to affect quality, frequent enough to matter for reliability.
The Credit System
Different models cost different amounts. Instead of per-message pricing, we use credits:
- **1 credit** ≈ one Haiku message (cheapest)
- **Pro messages** (Sonnet) ≈ 5 credits
- **Apex messages** (Opus) ≈ 25 credits
Free plan: 50 credits/month. That's roughly 50 simple edits or 10 full page generations. Enough to build your first website without paying anything.
Quality Verification
After every generation, we run a verification loop:
1. **Accessibility** — WCAG compliance, alt text, keyboard navigation
2. **Visual rendering** — does the page actually render correctly?
3. **Security** — no XSS vulnerabilities, no exposed secrets
4. **Performance** — image optimization, bundle size, load time
If verification fails, the system automatically attempts a fix before showing you the result. You only see the final, working output.
Why This Matters
Most users don't care which AI model generates their website. But they care about:
- **Speed** — Haiku responds in under a second for simple tasks
- **Reliability** — fallback chains mean 99.9%+ uptime for AI generation
- **Quality** — Opus handles the complex stuff that cheaper models can't
- **Cost** — you don't pay Opus prices for simple color changes
The routing system makes all of this invisible. You just type what you want and get a great result. That's the point.
[Try it yourself →](https://my.we.inc/projects/new)
We.Inc is an AI-powered website builder you can resell under your own brand. Launch a branded client dashboard, bill on Stripe Connect, and deliver AI-generated websites in minutes. White-label plans from $499/mo — no per-site fees.
Product
Who It's For
Features
Resources
Company
View Sitemap