RAG on Azure, AWS or Google Cloud: the real cost in production

A production RAG has four infrastructure items: search (the index), the model that writes the answer, embeddings at ingestion, and hosting for the application. Two of them decide almost everything.

1. Search. Billed by the hour, it costs the same whether your users ask ten questions or ten thousand. Billed by the query, it follows usage.
2. The model. It is billed per token, on input (retrieved passages plus the question) and output (the answer). The price depends on the model and on the deployment type you choose.

With our assumptions, embeddings cost little at ingestion: under $3 for 20,000 documents (details in the methodology). Application hosting, authentication and development are not priced here.

Sources: Azure Retail Prices API (francecentral), AWS Price List (eu-west-3, publications of 26 and 28 October 2026), Amazon OpenSearch Service, Amazon Bedrock and Google Cloud Agent Search pricing pages. Collected on 29 October 2026.

Two details change the calculation. On Azure, the same model has different prices depending on the meter: for gpt-4.1-mini in France Central, the API shows $0.40 / $1.60 for "global", $0.44 / $1.76 for "Data Zone" and $0.484 / $1.936 for "regional". On AWS, Nova 2.0 Lite costs $0.484 / $4.059 in the Paris region and $0.44 / $3.69 with "cross-region global" inference. The deployment type is therefore part of the price decision, and of the location decision too: read the provider's definition of each type before choosing.

Azure AI Search is billed per unit per hour, whether or not the service is used. Microsoft recommends at least two replicas for production workloads. Its service level agreement requires two replicas for read-only workloads and at least three for read-write workloads (queries and indexing). A Basic service with two replicas therefore costs 0.101 × 2 × 730 = $147.46 per month before the first question. An S1 service with three replicas: $735.84.

OpenSearch Serverless has two regimes, according to AWS's pricing page. For a classic collection, AWS bills at least 2 OCUs for the first collection in an account (one for indexing, one for search, each 0.5 × 2 for redundancy), i.e. 2 × 0.281 × 730 = $410.26 per month in Paris. A dev-test option without redundancy halves that minimum. With collection groups, AWS documentation states a configurable minimum of 0 OCUs: OCUs scale to zero after 10 minutes of inactivity. The bill then follows active hours, at the cost of a cold start that the documentation suggests avoiding by setting a minimum.

Google Agent Search and Bedrock Managed Knowledge Base are billed per query and per stored volume. With no usage, search costs almost nothing. Google also lists 10,000 free queries per account per month, presented as a trial, and 10 GiB of free storage per month.

In both scenarios priced below, this minimum weighs more than the model for Azure AI Search and for classic OpenSearch Serverless.

IgnitionAI estimate, based on the unit prices above and these assumptions: low scenario = 10 users, 2,000 documents (1 GB of raw data), 5,000 questions per month; realistic scenario = 50 users, 20,000 documents (10 GB), 25,000 questions per month; about 4,000 input tokens and 400 output tokens per question. One question = one search query. Free quotas not deducted.

Basis of the estimate: official prices of 29 October 2026, 730 hours per month, usage assumptions above. Excludes development, authentication, application hosting, logs, optional reranking and taxes. The models compared are not equivalent in quality: this table compares price structures, not answers.

Two honest readings of this table. First, at low volume, per-query services are the cheapest on infrastructure alone. Second, these amounts do not cover the heaviest item: building and maintaining the application around search. According to Google, Agent Search core answers do not include complex or long query handling or multimodality, which belong to a paid add-on.

A managed knowledge base retrieves passages and, depending on the service, writes an answer. An enterprise document assistant also needs:

• verifiable citations shown to the user;
• result filtering based on each person's access rights;
• a chat interface, authentication and account management;
• isolation per client if you serve several organisations;
• connectors that resynchronise sources;
• quality evaluation before every model or index change.

Each of these components has to be built, tested and maintained. That is the cost our build vs buy RAG guide sets against subscriptions. Our comparison pages for Azure, AWS and Google Cloud list the components to build on each cloud.

Basis of the grid: price structures collected on 29 October 2026 and the estimates above. Matching situations to approaches is our own reading.

Official prices (collected on 29 October 2026, HTTP 200)

• Azure Retail Prices API, region francecentral: Azure Cognitive Search (Basic, Standard S1), Foundry Models / Azure OpenAI (gpt-4.1-mini, text-embedding-3-small).
• Microsoft Learn, reliability in Azure AI Search (French edition): two replicas for the read-only SLA, three for read-write.
• AWS Price List, Amazon OpenSearch Service, eu-west-3, published 26 October 2026: SearchOCU and IndexingOCU.
• Amazon OpenSearch Service pricing page: 2-OCU minimum for the first classic collection, dev-test option, scale to zero after 10 minutes of inactivity.
• AWS, managing capacity limits for OpenSearch Serverless: 0-OCU minimum per collection group.
• AWS Price List, Amazon Bedrock, eu-west-3, published 28 October 2026: Nova 2.0 Lite, Titan Text Embeddings V2.
• Amazon Bedrock pricing page, Knowledge Bases section: Managed Knowledge Base. The page does not state the region for these prices.
• Google Cloud, Agent Search pricing: Standard and Enterprise editions, index storage, free quotas.

IgnitionAI estimates

• Scenarios and volumes: IgnitionAI usage assumptions (10 users, 2,000 documents, 5,000 questions per month; 50 users, 20,000 documents, 25,000 questions per month), the same as on our comparison pages.
• Average of 0.5 MB of raw data per document; about 4,000 input tokens and 400 output tokens per question; 5,000 tokens per document at ingestion.
• Embeddings at ingestion, with these assumptions: $0.22 (2,000 documents) and $2.20 (20,000 documents) with text-embedding-3-small Data Zone at $0.022 per million tokens; $0.30 and $3 with Titan Text Embeddings V2 at $0.03 per million. Embeddings are included in Managed Knowledge Base.
• 730 hours per month. Calculations rounded to the cent.

Limits

Cloud prices change often. This guide does not price development, maintenance, authentication, application hosting or taxes. It does not compare model quality. Check with each provider whether a service is available in a given region.

Last source review

29 October 2026.

To compare these amounts with a ready-to-use platform: IgnitionRAG offers the Solo (€59 excl. VAT/month) and Team (€249 excl. VAT/month) plans, with managed AI and credits included, and no key to provide. The trial includes 14 days and 100 credits, without a payment card. For a licensed deployment in your cloud, contact us from the pricing page.