Model Router (working name)
If model choice is a cost/quality tradeoff per task, why is anyone still routing everything to one model?
A smart API proxy for SMBs that automatically routes AI requests to the right model — balancing cost, speed, and quality. Users connect their API keys from multiple providers (Anthropic, OpenAI, Google, Mistral, etc.) and call one unified endpoint. The router classifies each request by task type and sends it to the optimal model. They get better results and lower bills without knowing which model ran.
The pain
SMBs using AI at scale — agencies, content teams, support ops — are either:
- Defaulting to one model for everything (overpaying for simple tasks, getting bad results on complex ones)
- Manually picking models per task (requires expertise they don't have)
How it works
- User connects their API keys (Anthropic, OpenAI, Google, Mistral — bring your own keys)
- They call one endpoint:
api.modelrouter.com/v1/chat - Router classifies the request: writing / summarization / analysis / coding / customer service / structured extraction / etc.
- Routes to the optimal model for that task type, given cost + quality tradeoffs
- Returns result. User sees a dashboard: cost saved, tasks routed, model breakdown.
Business model
- Small monthly subscription (access to the router, dashboard, routing rules)
- Usage markup on tokens processed through the platform (thin margin, e.g. 10-15% on top of provider cost)
- Power tier: custom routing rules, priority models, team seats
Who it's for
SMBs spending >$200/month on AI APIs who aren't deep enough to optimize themselves. Primary: marketing agencies, content teams, SaaS companies with AI features, small dev shops.
Differentiation vs existing tools
| Model Router | OpenRouter | LiteLLM | Martian | |
|---|---|---|---|---|
| Target user | SMB / non-technical | Developer | Developer | Enterprise |
| Routing logic | Auto (task classification) | Manual (user picks) | Manual / code | Auto |
| UI | Yes | Minimal | None | Yes |
| BYOK | Yes | Yes | Yes | No |
| SMB pricing | Yes | Usage only | Open source | Enterprise |
Riskiest assumption
SMBs who'd benefit most are already spending meaningfully on AI APIs AND willing to add a middleware layer between them and their providers. Security friction (sharing API keys with a third party) may block adoption. Prove: find 5 SMBs spending >$200/month on AI, ask if they'd pay $29/month + 10% markup for routing. If 3 say yes, build.
Open questions
- Name (working title: Model Router — needs something more brandable. Candidates: Relay, Conductor, Prism, Cascade)
- How does routing classification work? LLM-based classification (fast, adds latency + cost) or rule-based on request metadata?
- What happens when a user's preferred model is down? Fallback logic needed.
- Does the platform need to store API keys, or proxy them per-request? (Security architecture decision)
- Compliance: some SMBs can't have their data touch a third-party proxy (HIPAA, legal). Segment them out early.