What RAG search over your docs costs a month in API fees
Cheapest listed price for this workload
$14.25 a month
DeepSeek V4 Flash (0731) at DeepInfra. $0.29 per 1,000 requests.
You're answering questions over your own documents: retrieve the relevant chunks, then ask a model. At our example volume, the cheapest open model we track costs $14.25 a month in API fees, from the API prices we checked on 2026-11-02. Your bill scales with your volume, so the cost per 1,000 requests is the number to carry over.
Every model we track, cheapest first
Monthly API cost at the example volume, USD. Sorted by price only: we don't rank quality, so test the cheap ones on your own prompts.
| Model | Cheapest at | Monthly | Per 1,000 requests | Priciest listing | Providers |
|---|---|---|---|---|---|
| DeepSeek V4 Flash (0731) | DeepInfra | $14.25 | $0.29 | $29.25 | 2 |
| gpt-oss-20b | Groq | $18.75 | $0.38 | $18.75 | 1 |
| Llama 3.3 70B | DeepInfra | $24.00 | $0.48 | $221.00 | 2 |
| Qwen3 235B A22B Instruct 2507 | DeepInfra | $24.88 | $0.50 | $47.50 | 2 |
| gpt-oss-120b | Baseten | $26.25 | $0.53 | $37.50 | 4 |
| GLM-5.3 Flash | Baseten | $36.25 | $0.73 | $36.25 | 3 |
| DeepSeek V4.1 Flash | Baseten | $75.00 | $1.50 | $75.00 | 2 |
| Nemotron 3 Ultra | DeepInfra | $127.50 | $2.55 | $150.00 | 3 |
| GLM-5.3 | Baseten | $335.00 | $6.70 | $335.00 | 3 |
| Kimi K3 | DeepInfra | $748.13 | $14.96 | $787.50 | 4 |
The example volume: 50,000 questions a month, each sending about 4,000 input tokens (the question plus retrieved chunks) and getting about 250 output tokens back. Embedding costs aren't included. That is 200 M input and 12.5 M output tokens a month. Each model links to its page with every provider's input and output price and a link to that provider's own pricing page.
Self-hosting starts with $2,109.70 a month of rent
The cheapest NVIDIA H100 cloud VM we list is $2.89 per hour at RunPod, or $2,109.70 for 730 hours. The rent alone is about 148 times the cheapest API bill, so at this volume the API is cheaper; your traffic would have to grow by that much before one GPU's rent matched the API bill. That's rent only: it leaves out engineering time and idle hours, and whether one GPU can serve your traffic depends on the model, its size and your serving engine. Bigger models need several GPUs. Put your own throughput into the break-even calculator before you decide.