What will your RAG application cost?
See what indexing your documents, answering questions and running the application cost, and which change to test first.
RAG uses your documents to help an AI answer questions. The model is only one part of the cost.
Everything as a metered service: nothing to run, and a monthly minimum on the vector database.
- Embeddings: OpenAI text-embedding-3-small, $0.02 per 1M tokens
- Vector database: Pinecone Standard (serverless), $50.00 monthly minimum
- Answers: GPT-6.1 Terra, $1.00 in and $5.00 out per 1M tokens
- Application hosting: example $52.56/mo, a planning assumption
How often does the knowledge base change?Show
About 435 documents processed a month. Updates replace existing records; they do not add to storage.
How is content searched and passed to the model?Show
Each document becomes 5 chunks (the last 200 tokens), 2,200 tokens to embed with the overlap: 25,000 vectors in all.
Prompt per question: 2,730 tokens; answer 400 tokens.
Add rerankingShow
Reranking: a second step that sorts retrieved passages before they are sent to the language model.
Infrastructure and other costsShow
Pinecone Standard (serverless) is priced from its published rates. Choose a different vector database under “Change the models or use my prices”.
Change the models or use my pricesShow
Embedding: a numerical representation used to search your content. Your own prices are labeled as yours; the verified list is not changed.
Billed by storage, reads and writes, with a $50 monthly minimum that is billed instead of lower usage. Prices for AWS us-east-1.
What drives the bill
- Largest cost driver
- Answer generation (GPT-6.1 Terra), $71.94/mo (41%)
- First month (indexing plus one month)
- $174.74
- Fixed or minimum: $102.27
- Documents: $0.25
- Questions: $72.01
Fixed costs stay when traffic is low; document costs follow the knowledge base; question costs follow traffic.
A change worth testing
Each change keeps everything else as it is, and every affected cost is worked out again. Savings from separate changes do not add up: tick two or more to compare them together.
- Compare a lower-priced generation modelGeneration model: GPT-6.1 Terra to GPT-5.6 Luna; everything else as it is.
- Now
- $174.52/mo
- With this change
- $118.19/mo
- One-time cost to switch
- None
Potential cost difference under these assumptions: $56.33 a month lessA smaller model can answer less accurately or follow instructions less well. Run your own evaluation questions on both before switching.
- Smaller chunksChunk size: 500 to 250 tokens (overlap 50 to 25); everything else as it is.
- Now
- $174.52/mo
- With this change
- $159.65/mo
- One-time cost to switch
- $0.22
Potential cost difference under these assumptions: $14.87 a month lessEvery document is embedded and loaded again once: $1.71 of usage, but the record writes fall under the vector database's monthly minimum.
Shorter passages make each prompt smaller but create more vectors to store and search, and a passage can lose the context around it.
- Include fewer passages in each promptPassages in each answer: 5 to 3; everything else as it is.
- Now
- $174.52/mo
- With this change
- $161.14/mo
- One-time cost to switch
- None
Potential cost difference under these assumptions: $13.38 a month lessThe model sees less of your content, so check that answers to your own test questions still find what they need. Search, storage and query-embedding charges do not change.
Most of the vector database line is its $50.00 monthly minimum: usage is about $0.29. At this size, a database you already run may cost less. Choose it under “Change the models or use my prices” and enter what it adds.
Caching is not modeled here. For prompt-cache savings, use the LLM Token Cost Calculator.
Where the money goes each month
Open a line for its quantity, rate, formula and source.
Document updates and processing$0.02Verified price<1%Show
- Re-embedding changed documents$0.019
- Quantity
- 955,952 tokens
- Rate
- $0.02 per 1M tokens
- Formula
- 2% of 5,000 documents x 4.35 cycles a month = 434.5 documents x 2,200 tokens = 955,952 tokens; 955,952 / 1,000,000 x $0.02
- Source
- Verified price OpenAI list price, text-embedding-3-small, checked 30 Oct 2026
Vector database: Pinecone Standard (serverless)$50.00Verified price29%Show
- Vector storage$0.067
- Quantity
- 0.2032 GB-months
- Rate
- $0.33 per GB-month
- Formula
- 25,000 records x 8,128 bytes = 0.2032 GB; x $0.33
- Source
- Verified price Pinecone Standard list price, AWS us-east-1, checked 30 Oct 2026
- Searches (read units)$0.061
- Quantity
- 3,802 read units
- Rate
- $16.00 per 1M read units
- Formula
- 15,208 searches x 0.25 read units each (1 read unit per GB of namespace, at least 0.25 a query) = 3,802 read units; / 1,000,000 x $16
- Source
- Verified price Pinecone Standard list price, AWS us-east-1, checked 30 Oct 2026
- Record writes for updates and new documents$0.159
- Quantity
- 35,327 write units
- Rate
- $4.50 per 1M write units
- Formula
- 0 new and 2,172.6 overwritten records of 8,128 bytes (1 write unit per KB of request, at least 5 a request, plus 1 per KB of a record overwritten; 100 records a request assumed) = 35,326.8 write units; / 1,000,000 x $4.5
- Source
- Verified price Pinecone Standard list price, AWS us-east-1, checked 30 Oct 2026
- Monthly minimum adjustment (Standard plan)$49.71
- Quantity
- 1 month
- Rate
- $50.00 minimum per month
- Formula
- Usage $0.29 is under the $50.00 minimum, which is billed instead: + $49.71
- Source
- Verified price Pinecone Standard plan minimum, dedicated account assumed, checked 30 Oct 2026
Question embeddings$0.01Verified price<1%Show
- Question embeddings$0.00912
- Quantity
- 456,250 tokens
- Rate
- $0.02 per 1M tokens
- Formula
- 15,208.3 questions x 1 search x 30 query tokens = 456,250 tokens; 456,250 / 1,000,000 x $0.02
- Source
- Verified price OpenAI list price, text-embedding-3-small, checked 30 Oct 2026
Answer generation (GPT-6.1 Terra)$71.94Verified price41%Show
- Prompt tokens$41.52
- Quantity
- 41,518,750 tokens
- Rate
- $1.00 per 1M tokens
- Formula
- 15,208.3 questions x (500 instructions + 30 question + 0 history + 0 other + 5 passages x 440 = 2,730 tokens)
- Source
- Verified price GPT-6.1 Terra list price, checked 30 Oct 2026
- Answer tokens$30.42
- Quantity
- 6,083,333 tokens
- Rate
- $5.00 per 1M tokens
- Formula
- 15,208.3 questions x 400 answer tokens
- Source
- Verified price GPT-6.1 Terra list price, checked 30 Oct 2026
Application hosting$52.56Planning assumption30%Show
- Application hosting (example)$52.56
- Quantity
- 1 month
- Rate
- $52.56 per month
- Formula
- 730 hours x 3,600 seconds x (1 vCPU x $0.000018 + 1 GiB x $0.000002) per second, instance-based billing, free tier not applied
- Source
- Planning assumption one Cloud Run instance kept running (1 vCPU, 1 GiB, Iowa). An example, not a quote: replace it with your own figure., checked 1 Nov 2026
Initial indexing, one time: $1.13
- Document embeddings$0.220
- Quantity
- 11,000,000 tokens
- Rate
- $0.02 per 1M tokens
- Formula
- 5,000 documents x 2,200 tokens (5 chunks, overlap included) = 11,000,000 tokens; 11,000,000 / 1,000,000 x $0.02
- Source
- Verified price OpenAI list price, text-embedding-3-small, checked 30 Oct 2026
- Record writes for the first load$0.915
- Quantity
- 203,250 write units
- Rate
- $4.50 per 1M write units
- Formula
- 25,000 new and 0 overwritten records of 8,128 bytes (1 write unit per KB of request, at least 5 a request, plus 1 per KB of a record overwritten; 100 records a request assumed) = 203,250 write units; / 1,000,000 x $4.5
- Source
- Verified price Pinecone Standard list price, AWS us-east-1, checked 30 Oct 2026
If questions grow
Same documents, more questions. Each row is priced again in full, minimums included.
| Questions a day | Monthly | Per 1,000 questions | Note |
|---|---|---|---|
| 500 | $174.52 | $11.48 | Today |
| 2,500 (5x) | $462.30 | $6.08 | Capacity validation required |
| 5,000 (10x) | $822.02 | $5.41 | Capacity validation required |
Hosting and other fixed amounts are held at today’s figures; check they can serve the higher load. Corpus growth is a separate setting.
What this estimate leaves outShow
- Prompt caching, semantic answer caches and batch discounts: every question is priced at the standard rate.
- Query rewriting, agent loops, tool calls and follow-up model calls: one generation call per question.
- GraphRAG, images, audio and scanned-page understanding.
- Fine-tuning or training models.
- Engineering and operations time, support plans, taxes and negotiated discounts.
- Deleting stale records when a document shrinks: updates are priced as overwrites of the same records.
- Text extraction, OCR and parsing: not entered. Scanned PDFs usually cost more to process than clean text.
- Storage for the source documents: not entered.
- Logs and monitoring: not entered.
- Data transfer between clouds or regions: not entered.
- Pinecone backups, restores, imports, inference and support add-ons.
- Pinecone Starter (free) and Builder plans: the estimate assumes the paid Standard plan on its own account.
- Pinecone write units are published as $4 to $4.50 per million by cloud and region; the upper end is used.
What this model assumes
Show
- One pipeline, priced stage by stage. Each question is embedded once per search, searched once, optionally reranked, and answered by one generation call. No query rewriting, agent loops, semantic cache or prompt-cache discount.
- Chunking. Each document is split on its own into windows of the chunk size, each starting the chunk size minus the overlap after the last; the final, shorter chunk is billed at its own length. Real splitters, document lengths and tokenizers change the count.
- Updates replace records. Reprocessed documents overwrite their records and do not add storage. Only new documents grow the knowledge base, priced on the month’s average. Changing the embedding model or dimension means embedding and loading every document again.
- Minimums are floors. A plan minimum is billed instead of lower usage, not on top of it, and only once a month: the first month puts indexing and the month’s usage under the same floor. One account per application is assumed.
- Price states. Verified prices come from each provider’s published rates; your prices and planning assumptions are labeled as such; a missing price makes the estimate incomplete, never zero. No compliance, industry or “enterprise” multiplier is applied.
- Not a capacity plan. Nothing here guarantees throughput, latency or answer quality. A forecast helps plan spending. Actual usage and invoices validate it.
- OpenAI text-embedding-3-small $0.02 per 1M tokens, checked 30 Oct 2026.
- OpenAI text-embedding-3-large $0.13 per 1M tokens, checked 30 Oct 2026.
- Google gemini-embedding-001 $0.15 per 1M tokens, checked 30 Oct 2026.
- Voyage AI voyage-4 $0.06 per 1M tokens, checked 30 Oct 2026.
- Voyage AI voyage-4-lite $0.02 per 1M tokens, checked 30 Oct 2026.
- Amazon Bedrock Titan Text Embeddings V2 $0.02 per 1M tokens, checked 30 Oct 2026.
- Voyage AI rerank-3 $0.05 per 1M tokens (query tokens x candidates plus every candidate's tokens, as Voyage counts them), checked 30 Oct 2026.
- Voyage AI rerank-3-lite $0.02 per 1M tokens (query tokens x candidates plus every candidate's tokens, as Voyage counts them), checked 30 Oct 2026.
- Pinecone Standard (serverless) (AWS us-east-1): storage $0.33 per GB-month, reads $16.00 and writes $4.50 per 1M units, $50.00 monthly minimum, checked 30 Oct 2026.
- Generation models from this site’s LLM price catalog, each with its own check date in the breakdown.
- Application hosting example: one Cloud Run instance kept running (1 vCPU, 1 GiB, Iowa), from Cloud Run list prices checked 1 Nov 2026. A planning assumption.
Prices verified 30 Oct 2026. List prices in USD for Pinecone on AWS us-east-1; models not region-specific, plus the rates you enter.
Want your real AI spend, not an estimate?
Finitizer Tokenomics: Reads your AI bills and exports and shows what changed, who spent it, and what to fix. No proxy, no code change. Founding customers list now open.
Questions
What does a RAG application cost to run?
It depends on four things: how many questions it answers, how much text goes into each prompt, the vector database plan and the hosting. In the example on this page, 5,000 documents and 500 questions a day on the managed, pay-per-use configuration come to about $175 a month, and answer generation is the largest part.
Is the language model the main cost of RAG?
Often, but not always. With few questions, the vector database minimum and the hosting can cost more than the model. With many questions or long prompts, generation dominates. The calculator finds the largest cost driver from your own figures rather than assuming one.
How much does it cost to embed my documents?
Document embeddings are usually small. At $0.02 per million tokens, the example's 5,000 documents of 2,000 tokens, split into 500-token chunks with 50 tokens of overlap, are 11 million tokens: $0.22 to embed once. Re-embedding only the documents that change keeps the monthly cost low.
Does a vector database minimum add to my usage?
No. Pinecone's Standard plan has a $50 monthly minimum that is billed instead of lower usage, not on top of it. If a month's usage is $12, the bill is $50; at $80 of usage it is $80. The first month puts the indexing writes under the same floor.
What does this calculator leave out?
Prompt caching, semantic caches, query rewriting, agent loops, scanned-document processing unless you enter it, engineering time, taxes and negotiated discounts. Each exclusion is listed on the page rather than hidden in a multiplier.
Related tools
LLM Token Cost Calculator
Compare monthly AI spend on Claude, GPT, Gemini and more before the invoice arrives.
OpenAI Agent Cost Calculator
Agents make many model calls per task. Price the whole loop, not one request.
OpenCloud Run Cost Calculator
Two billing models, and how busy your instances are picks the winner.
OpenMore tools for: AI and model usage, Storage and lifecycle
A forecast helps plan spending. Actual usage and invoices validate it.
Finitizer tracks real token consumption per feature and team across your AI providers, so you can compare this estimate with what your RAG application actually used.
See Finitizer Tokenomics