Real-World Token Economics for RAG Search & Vector Ingestion
RAG workloads are heavily asymmetrical: 95% of token volume is input (retrieved document chunks), with very little generated output.
Typical monthly workload
- Input tokens
- 15,000,000
- Output tokens
- 500,000
- Requests / day
- 2,000
Illustrative estimate. Adjust the sliders below to match your own volume.
Recommended setup
Embed initial chunks with OpenAI embedding ($0.02/M) or bypass chunking entirely by passing full PDF books directly into Gemini 1.5 Pro.
Monthly cost by model for this workload
| Model | Input / 1M | Output / 1M | Est. monthly cost |
|---|---|---|---|
| Gemini 1.5 Flash | $0.075 | $0.30 | $1.27 |
| DeepSeek V3 / V2.5 | $0.140 | $0.28 | $2.24 |
| GPT-4o Mini | $0.150 | $0.60 | $2.55 |
| Llama 3.3 70B (Groq/Together) | $0.590 | $0.79 | $9.24 |
| Claude 3.5 Haiku | $0.800 | $4.00 | $14.00 |
| o1-mini | $1.100 | $4.40 | $18.70 |
| Gemini 1.5 Pro | $1.250 | $5.00 | $21.25 |
| Mistral Large 2 | $2.000 | $6.00 | $33.00 |
| GPT-4o (Omni) | $2.500 | $10.00 | $42.50 |
| Claude 3.5 Sonnet | $3.000 | $15.00 | $52.50 |
| o1 (Reasoning) | $15.000 | $60.00 | $255.00 |
Computed from list prices and the workload above (15,000,000 input and 500,000 output tokens per month), with no caching or batch discounts. Real bills vary.
How to cut this cost
- Bypass Vector Databases with 2M Context Models
For small enterprise intranets (<200 documents), passing documents directly into Gemini 1.5 Pro eliminates vector DB infrastructure costs completely.
- Apply Reranking Before Synthesis
Use Cohere Rerank to filter 50 retrieved chunks down to the top 3 most relevant before sending to the LLM, cutting input tokens by 80%.
Simulate your own spend
Quick Workload Presets
Pick a ready-made scenario or customize your token volume below.
Token & Usage Parameters
By switching or routing background tasks from GPT-4o to Gemini 1.5 Flash, your team saves up to $2,270 annually.
Lowest cost per token, high-frequency tasks, real-time audio/video
State-of-the-art open economics, high-performance general intelligence
High-volume text processing, fast chatbots, classification
Open-weights privacy, self-hosting or ultra-cheap hosted inferencing
Rapid code completion, customer service, structured data extraction
Fast STEM reasoning, competitive programming at lower cost
Massive context (2M tokens), full video/audio analysis, large codebases
Multilingual tasks, European compliance, enterprise code generation
Complex multimodal reasoning, coding, and production apps
Software engineering, nuanced writing, complex agentic workflows
Advanced mathematics, algorithmic coding, PhD-level scientific logic
Frequently asked questions
How much does it cost to index 1,000 PDF pages?
1,000 standard PDF pages contain roughly 500,000 words (~650,000 tokens). Embedding this volume with OpenAI text-embedding-3-small costs approximately $0.013.