Companies may be wasting money on LLM calls.

pruthvep — Fri, 21 Aug 2026 23:11:05 +0000

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?

👉 Try AI Router Live

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.

DEV Community: pruthvep