LLM Model Routing Cost Optimizer
Calculate real savings from routing tasks to cheaper models. Unlike competitors, this calculator includes router overhead and misroute penalties — the costs that eat into theoretical savings.
| Component | Requests | Model | Cost | Notes |
|---|---|---|---|---|
| Simple tasks | 500,000 | GPT-5 mini | $587.50 | 50% of volume |
| Moderate tasks | 300,000 | Claude Sonnet 5 | $3,150.00 | 30% of volume |
| Complex tasks | 200,000 | Claude Opus 5 | $3,500.00 | 20% of volume |
| Router overhead | 1,000,000 | — | $52.50 | 200 tokens each |
| Misroute penalty | 50,000 | — | $800.00 | 5% retry rate |
| Total routed | — | — | $8,090.00 | vs $9,750.00 baseline |
OpenRouter ↗ — managed model routing with automatic fallbacks. Compare all models on our model comparison table.
Prices verified 2026-08-20 · Formula documentation
How to Use
Set your task mix
What percent of your requests are simple (classification, extraction), moderate (summarization), or complex (reasoning, code)? The percentages must sum to 100%.
Choose models per tier
Pick the cheapest appropriate model for simple tasks, a mid-tier model for moderate, and your best model for complex. Set your current baseline model.
Set volume and router config
Enter total monthly requests, average token counts. If using a router model, enable it to add overhead and misroute costs to the estimate.
See honest savings
Unlike other calculators, this one includes router overhead and misroute penalties — the costs competitors ignore. Your net savings is what you'll actually see in billing.
The honest routing formula
Every routing calculator shows you gross savings. Few show net savings:
Baseline = requests × (in_tokens×P_in_base + out_tokens×P_out_base) / 1M Routed = Σ_tiers(volume_t × (in×P_in_t + out×P_out_t) / 1M) Router overhead = requests × (router_tokens+10) / 1M × P_in_cheapest Misroute penalty = requests × misroute_rate × retry_cost Net savings = Baseline − (Routed + Router_overhead + Misroute_penalty)
At 5% misroute rate and 200-token router prompts, overhead typically reduces gross savings by 15–25%. At 15% misroute rate it can reduce savings by 50%. Both numbers are in this calculator.
How to Cut This Cost
Route 70–80% of simple queries to a cheap model (GPT-5 mini / Haiku 4.5). Reserve the capable model for complex, high-stakes tasks.
Reduce misroute rate by improving your router — even a basic classifier (BM25 or small embedding model) beats random routing by 30–50%.
Add a Batch API tier for non-realtime simple tasks. Combine routing + batch for compounding savings.
Lower router token cost by using a short, schema-driven classification prompt rather than a verbose routing prompt.
FAQ
Have more questions? Contact us
Last verified: 2026-08-20 · methodology · data sources