When Nadir helps most
High volume, mixed complexity: If you're making 1000+ API calls/day and many are simple lookups, formatting, or basic Q&A, routing has more room to help. Nadir's older 60% / 98% RouterBench result is a reference-assisted research ceiling on 11,420 triples. Do not forecast from it; start with your declared baseline and outcome-labelled shadow traffic.
Agentic workflows: Claude Code, Cursor, Aider - these tools make hundreds of calls per session. Many are simple (read file, check status). Routing + context optimization can cut costs 50%+.
BYOK setups: When you bring your own API keys, you pay provider prices directly with no markup on the model. Nadir's only charge is a flat $0.002 per request.
When to stick with one model
All-complex workloads: If every prompt is a complex reasoning task (legal analysis, medical diagnosis), routing won't help. Everything stays on premium anyway.
Latency-critical paths: The 50-100ms classification overhead matters if you're in a real-time pipeline where every millisecond counts. For most use cases, this is negligible.
Small volume: If you're spending <$50/month on LLM APIs, the operational overhead of running Nadir isn't worth the savings.
The math
| Monthly spend | Prompt mix | Savings |
|---|---|---|
| $500 | 60% simple, 30% medium, 10% complex | ~$180 |
| $2,000 | 40% simple, 40% medium, 20% complex | ~$640 |
| $10,000 | 30% simple, 40% medium, 30% complex | ~$2,800 |
Nadir takes no fee on savings. Its only charge is a flat $0.002 per request, so subtract your monthly request count times $0.002 from the saving.
Bottom line
If your workload has a mix of simple and complex prompts, Nadir pays for itself from day one. If everything you do needs premium reasoning, save yourself the setup.