₩ Cost & build-vs-buy
Korean LLM cost calculator — API vs self-host
Estimate monthly cost by call volume, compare API per-token pricing against self-hosting open-weight models, and see where the crossover is for your workload.
API models — estimated monthly cost
| Model | $/1M in | $/1M out | Monthly |
|---|---|---|---|
| Solar Pro 3✓검증cheapest | ₩63,000 / $45.00 | ||
| Gemini (참고)참고 | ₩525,000 / $375 | ||
| GPT (참고)참고 | ₩1,050,000 / $750 | ||
| Claude (Sonnet급 참고)참고 | ₩1,470,000 / $1,050 |
Self-host (open weights) — infra cost, volume-independent
- ▪A.X 4.0 (72B) — Open weights · self-host GPU infra (fixed, volume-independent)
- ▪EXAONE 4.0 32B — Open weights (NC license) · self-host infra
- ▪Trillion Tri-7B — Lightweight open weights · minimal infra
Solar Pro 3 price verified (Upstage). Others are editable reference defaults — confirm on official pages. HyperCLOVA X is console-priced (enter your rate). Korea-tuned models can need ~30% fewer Korean tokens (A.X reports +33% vs GPT-4o) — adjust input/output tokens accordingly. Self-host trades per-token cost for fixed GPU infra; the crossover depends on volume.
How to read it
- ›API cost scales with volume; self-host is a fixed GPU infra cost — high volume favors self-host, low/spiky volume favors API.
- ›Korean token efficiency matters: Korea-tuned models can use ~30% fewer Korean tokens, lowering real cost below the headline per-token price.
- ›Open-weight commercial terms differ (A.X: Qwen License MAU<100M; EXAONE: non-commercial) — factor licensing into 'self-host'.
Get the real number for your prompts
A 48h diagnostic measures these on your own 20 prompts and returns a cost + buy/wait conclusion. ₩490,000 · tax invoice.