Introduction
Several CTOs end up spending 1.5 to 2 times their initial AWS Bedrock estimates. But the overspend doesn't come from hidden fees; it comes from AWS Bedrock costs that are difficult to calculate upfront because the official AWS pricing page only shows per-token numbers. The actual bill tells a different story.
On-demand inference, provisioned throughput, batch processing, Knowledge Bases, Guardrails, and adjacent AWS services all contribute independently to the cost. Most teams only connect the dots after the fact.
This guide breaks down each of these cost components into 6 steps so CTOs and cloud teams can calculate AWS Bedrock costs accurately and set a realistic budget from the start.
Step 1: Pick the Right AWS Bedrock Pricing Model
Bedrock offers four pricing tiers, and each one has different economics. Picking the right one is the first decision in any cost calculation.
On-Demand charges per million tokens with no commitment. AWS enforces rate limits, so traffic spikes can trigger throttling, but there's no minimum spend. Right for prototypes, variable workloads, and anything early-stage.
Batch Inference takes a JSONL prompt file from S3, processes it asynchronously within 24 hours, and returns results to S3. AWS prices Batch at 50% below on-demand rates for Anthropic, Meta, Mistral, and Amazon models. Document pipelines, nightly embedding backfills, evaluation runs, anything asynchronous should default here.
Provisioned Throughput reserves dedicated model capacity in Model Units (MUs), billed hourly whether used or not. Discounts of 15% to 40% apply on 1-month or 6-month terms. At sustained utilization above 80%, it beats on-demand. Below 60%, idle capacity erases the discount. It's also the only way to run fine-tuned or custom-imported models on Bedrock.
Cross-Region Inference routes requests to another region during capacity constraints, billed at the source region rate with no surcharge. It's a resilience mechanism, not a cost lever.
Getting this wrong means building a cost model on the wrong foundation. A team estimating on-demand who later shifts to Provisioned Throughput at scale is working with two very different numbers.
With the pricing tier decided, the next step is calculating base token costs.
Step 2: Calculate Base Token Costs
Across almost every model on Bedrock, output tokens cost three to five times more than input tokens. Claude Sonnet 4.6 runs $3.00 per million input versus $15.00 per million output, a 5x gap. For chatbots where responses run long, that gap matters more than the headline input rate.
Here is the formula to calculate the base token costs:
Monthly token cost = (monthly input tokens / 1,000,000) x input rate + (monthly output tokens / 1,000,000) x output rate
A token is roughly 4 characters or three-quarters of a word. System prompts, conversation history, retrieved context, and tool results all count as input. The model's response is output.
Current 2026 on-demand rates in us-east-1:
- Claude Opus 4.8: $5.00 input / $25.00 output per million tokens
- Claude Sonnet 4.6: $3.00 input / $15.00 output per million tokens
- Claude Haiku 4.5: $1.00 input / $5.00 output per million tokens
- GPT 5.4: $2.50 input / $15.00 output per million tokens
- Amazon Nova Lite: $0.06 input / $0.24 output per million tokens
Amazon Nova Micro: $0.035 input / $0.14 output per million tokens
Nova Micro at $0.035 per million output tokens is roughly 143 times cheaper than Claude Opus 4.8. For classification, routing decisions, or structured extraction, defaulting to a frontier model is one of the most expensive habits in production AI.
Example: A support chatbot on Claude Haiku 4.5 at 50,000 monthly queries, averaging 800 input and 400 output tokens:
- Input: 40M tokens x $1.00/M = $40
- Output: 20M tokens x $5.00/M = $100
- Base total: $140/month
That's straight inference only, no agents, no Knowledge Bases, no Guardrails.
Step 3: Apply the Agent Token Multiplier
The base formula above works for simple inference, but it breaks down the moment agents enter the picture. Agentic workloads consume far more tokens than a prompt-and-response estimate suggests. A single user query routed through Bedrock Agents triggers multiple internal model calls: reasoning about which tool to use, calling it, interpreting the result, deciding whether to call another, and generating the final response. Each step bills at the same per-million rate.
Multiply base inference costs by a workload multiplier:
- Simple chatbot: 1x
- RAG assistant: 2x to 4x (retrieved context stuffed into every input)
- Agentic workflows with tool calls: 5x to 10x
An agentic workflow handling 5,000 monthly runs on Claude Sonnet 4.6 with 2,000 visible input tokens and 500 output tokens shouldn't be estimated against the base formula. At a 6x multiplier, actual token consumption is 12,000 input and 3,000 output per run, shifting the monthly estimate from roughly $190 to over $1,100 before any other charges.
For teams running Bedrock AgentCore, the billing layers multiply further: Runtime charges per-second vCPU and GB-hours, Gateway charges per MCP operation, Memory charges per event and record, Identity charges per OAuth token, and Policy charges $0.000025 per authorization request. Each is billed independently.
Step 4: Factor in Knowledge Base and Vector Store Infrastructure
The multiplier in Step 3 covers compute, but teams running Retrieval-Augmented Generation face a separate cost layer entirely: the vector store. Knowledge Bases for Amazon Bedrock carry no platform fee, but the storage system underneath them does, and it's often the charge teams miss until the first invoice arrives.
The default vector store is Amazon OpenSearch Serverless, which requires a minimum of 2 OpenSearch Compute Units at $0.24 per OCU per hour. That works out to roughly $345 per month with zero query traffic. A team running 100 monthly Knowledge Base queries pays the same vector store floor as one running 100,000.
Since December 2025, Amazon S3 Vectors offers an alternative priced up to 90% cheaper with sub-second retrieval at scale. For any new Knowledge Base without a specific OpenSearch dependency, S3 Vectors is the lower-cost default.
Additional retrieval charges to include:
- Embedding model inference (input tokens only, at the embedding model's rate)
- Bedrock Data Automation at $0.010 per page for document parsing during ingestion
- Amazon Rerank 1.0 at $1.00 per 1,000 queries for deployments using semantic reranking
A mid-scale RAG assistant at 50,000 monthly queries typically runs $1,500 to $3,500 per month on inference and retrieval combined, before the vector store floor.
Step 5: Include Fine-Tuning, Guardrails, and Adjacent AWS Charges
Vector stores aren't the only cost sitting outside core inference. Fine-tuning, Guardrails, and a few adjacent AWS services also add up, and most cost estimates leave them out entirely.
Fine-tuning adds three separate cost lines: training charges per token (dataset size x epochs x per-token rate), monthly model storage at $0.02 to $0.10 per GB, and Provisioned Throughput inference billed hourly since custom models have no on-demand option. Teams that plan these charges upfront as part of a broader AWS migration services engagement typically land 20 to 25% closer to actual spend than teams who discover them post-launch.
Guardrails charges per text unit (1,000 characters), with content filters and denied topics each at $0.15 per 1,000 text units. At 1,000 queries per hour with Guardrails on both input and output, it's a real monthly line item.
On top of all of these, some adjacent AWS services round up the bill in ways most Bedrock estimates miss entirely:
- CloudWatch Logs at $0.50 per GB ingested. Verbose system prompts logged at volume add up fast.
- S3 for batch artifacts and Knowledge Base document storage
- KMS per API call for encryption at rest
- VPC endpoints billed hourly for private connectivity
- Data transfer between regions at standard EC2 rates, including for AgentCore Runtime, Gateway, and Code Interpreter traffic as of November 2025
On production workloads, these adjacent charges typically run 20% to 35% on top of inference.
For teams still planning a move to AWS or recently migrated, an AWS migration services provider can factor these charges into the cost plan from the start, instead of leaving them to surface after the workload goes live.
Step 6: Build the Full Monthly Estimate
With every layer covered, from base inference through fine-tuning and adjacent charges, the last step is adding them together into one number. A complete AWS Bedrock costs estimate includes:
- Base inference (token formula multiplied by the workload multiplier)
- Vector store floor (OpenSearch Serverless at roughly $345/month, or S3 Vectors significantly lower)
- Embedding model inference and reranking
- Guardrails, billed per text unit
- AgentCore sub-components (Runtime, Gateway, Memory, Identity, Policy), if applicable
- CloudWatch logging
- S3, KMS, VPC endpoints, and data transfer
For most production generative AI workloads, the honest total lands between $1,500 and $8,000 per month. Enterprise deployments running multiple agentic workflows against large Knowledge Bases regularly exceed $20,000 per month. None of that is surprising when the full picture is built before architecture decisions are made. It only catches teams off guard when the estimate stops at token rates.
Running this manually across seven or eight variables gets unwieldy fast, especially when modeling more than one workload scenario. For teams who'd rather skip this manual process, Bacancy's AWS Bedrock pricing calculator handles the same inputs covered above, including the agent multiplier and vector store baseline cost, and returns a monthly estimate in a few clicks.
Conclusion
To calculate AWS Bedrock costs, you need to add these six things: base inference, the agent multiplier, vector store infrastructure, fine-tuning, Guardrails, and the adjacent AWS services running alongside them. Skip one, and the bill three months from now won't match what was budgeted.
Teams that want an outside read before committing to an architecture can can leverage AWS consulting services to review the workload, recommend the right pricing model, and flag where costs are likely to run over budget.
Top comments (0)