Companies may be wasting money on LLM calls.
Many AI applications send every prompt to the same powerful model — whether it's a simple question or complex reasoning.
But does every prompt really need the most expensive model?
I built AI Router to test that idea.
How it works
Instead of:
Every Prompt → One Model
AI Router does:
Prompt → Classification → Difficulty → Routing Policy → Best Model
It benchmarks OpenAI, Gemini, and Claude on both quality and API cost, then routes prompts based on those results.
Initial Results
| Model | Quality | Avg. Cost |
|---|---|---|
| OpenAI | 96.90 | $0.0029509 |
| Gemini | 95.75 | $0.0012094 |
| Claude | 93.35 | $0.0032733 |
The interesting result:
Gemini was only 1.15 quality points below OpenAI, while its measured average request cost was ~59% lower.
This doesn't mean Gemini is always better. It shows why using the same model for every workload may not be the most cost-efficient approach.
Try It Live
Want to see how the router handles your prompt?
Enter a prompt and see how AI Router classifies the request and selects a model based on its routing policy.
👉 View the Source Code on GitHub
What's Next?
This is still an MVP with a small benchmark, so these results shouldn't be treated as universal model rankings.
Next, I'm testing 100–500+ prompts to compare:
Always use the strongest model
vs.
Dynamically route to the best quality/cost model
I'll measure:
Cost • Quality • Latency • Reliability
The question I'm trying to answer is:
How much can intelligent LLM routing reduce cost without significantly sacrificing quality?
Tech Stack
Built with Go, Gin, OpenAI, Gemini, Claude, Next.js, and TypeScript.
Open to Opportunities
I'm currently exploring AI/LLM Engineering, Backend Engineering, and AI Infrastructure opportunities where I can work on problems involving model routing, RAG, agents, and production AI systems.
If your team is working on similar problems, I'd love to connect.
I'm also open to feedback and conversations with engineers building production LLM systems.
Top comments (0)