Send every request to a frontier model and you overpay for easy queries. Split traffic across light, general, and complex tiers to see the monthly savings — net of router overhead.
100% free · no signup · browser-local
Share on XSet the share of requests sent to each tier. Percentages are normalized to 100%.
Intent detection, routing, extraction, simple FAQ answers.
Standard chat, summarization, and drafting.
Multi-step analysis, code, and high-stakes answers.
All frontier
$4,600.00
Smart routed
$905.50
Monthly savings
$3,694.50
Savings rate
80.3%
Per request (naive)
$0.00
Per request (routed)
$0.00
Traffic allocation, per-request cost, and monthly spend for each tier.
| Tier | Model | Share | Requests | Cost / req | Monthly |
|---|---|---|---|---|---|
| Light / classification | DeepSeek Flash | 55% | 550K | $0.00 | $165.00 |
| General support | GPT-6 Luna | 35% | 350K | $0.00 | $80.50 |
| Complex reasoning | Claude Sonnet 5.5 | 10% | 100K | $0.00 | $460.00 |
High-volume traffic with a predictable long tail of easy requests. The more skew toward simple queries, the bigger the savings.
A classifier or gateway call per request. Cheap relative to frontier inference, but it must be measured, not assumed free.
Always evaluate escalated answers against the frontier baseline. Savings are only real if quality holds on your task.
An intent classifier or small model labels each request as light, general, or complex. This is the router overhead line in the calculator.
Light requests hit a sub-$0.20 model, general requests hit a mid-tier model, and only hard requests reach the frontier tier.
Sample escalated and non-escalated answers. If the cheap tier fails, the retry cost can erase the routing savings.
Routing assumes the same average token counts for every tier and perfect tier selection. Real classifiers make mistakes, so validate quality before banking the savings.
Model multi-step agents with tool calls, retries, and monthly run volume.
Calculate GGUF weights + KV cache, match RTX / Apple Silicon / Strix Halo rigs, and compare Ollama, llama.cpp, MLX & vLLM.
Find the break-even volume where fine-tuning beats long few-shot prompts.
We'll email you when model prices change or new tools launch. Unsubscribe anytime.