Provider mix hides unit cost
Teams compare model cards but rarely track cost per completed workflow across Bedrock model choices.
AWS Bedrock Cost Review
Last reviewed Reduce AWS Bedrock costs before scaling agents and RAG.
Bedrock cost issues often come from provider mix, model choice, retrieval overhead, guardrail patterns, agent tools, and retries. NavyaAI maps Bedrock spend to completed workflow cost so teams can optimize before expanding usage.
50-500x
agentic usage multiplier
72%
of AI spend hides outside inference
42%
cost cut in a recent audit
RAG and agents multiply calls
Retrieval, reranking, guardrails, tool use, and retries create hidden cost outside the primary model call.
Architecture changes before measurement
Teams add more services before separating latency, quality, and token-volume problems.
Deep Dive
What one RAG query actually costs on Bedrock
A single answered question in a production Bedrock RAG system is never one model call. It is an embedding call for the query, a vector search, often a reranking pass, a generation call carrying retrieved context, and one or more guardrail evaluations on the way in and out. Add an agent layer and each tool decision is another billable invocation.
This is why teams that compare Bedrock model prices per million tokens still get surprised by the invoice: the meaningful unit is cost per answered query, and it is dominated by the call multiplier, not the per-token rate. The audit instruments that pipeline end to end before recommending any model or provider change.
Model mix and routing across the Bedrock catalog
Bedrock's strength — many models behind one API — is also its main cost trap. Without a routing policy, traffic settles on whichever frontier model the team trusted during development, and routine tasks ride along at premium rates. The catalog rewards the opposite: route extraction, classification, and formatting to lighter models, and reserve frontier capacity for the reasoning steps that measurably need it.
Guardrails follow the same logic. Safety checks priced per evaluation get repeated at multiple pipeline stages by default; scoping which checks run where, and batching them, trims cost without weakening the controls that matter.
When steady Bedrock volume justifies the break-even question
For predictable, high-volume workloads the comparison eventually extends beyond the Bedrock catalog to private serving. Our Token Tax benchmark measured optimized self-hosted Llama 3 70B at roughly $0.47 per million tokens — INT8 quantization and KV-cache pruning, 2.3x throughput on half the GPUs — against $1.80-$2.50 per million for frontier APIs at testing time.
That spread only pays off with high utilization and real operations capacity, which is why the audit produces break-even math from your traffic shape before any infrastructure recommendation.
Audit Focus
What we inspect before prescribing a platform change.
The first pass is designed to identify the smallest useful intervention: routing, caching, prompt control, serving tuning, or a deeper break-even audit.
Bedrock model selection and task routing
RAG retrieval, reranking, and context-size controls
Agent tool-use, retry, and guardrail cost patterns
Cost per completed user action
Bedrock vs Vertex/OpenAI/private route comparison
Bedrock cost review map — see the full map
The audit checks whether Bedrock cost pressure is pricing, workflow design, or orchestration overhead.
| Signal | Likely leak | Audit question |
|---|---|---|
| Multiple models | No routing policy by task class | Which tasks need the strongest model? |
| RAG chains | Context and rerank cost compounds | How many calls answer one user query? |
| Agent tools | Tool loops continue after enough evidence | Where should the loop stop? |
| Guardrails | Safety checks are repeated unnecessarily | Which checks can be batched or scoped? |
| Steady traffic | No private break-even model | Does cloud GPU or on-prem serving win? |
Worked Example
Cost stack of one answered RAG query
An illustrative call-by-call breakdown of a single user question in a Bedrock RAG pipeline with guardrails — the multiplier, not the token price, drives the bill.
| Pipeline step | Billable calls per query | Audit question |
|---|---|---|
| Query embedding | 1 embedding call | Can frequent queries be cached? |
| Retrieval + reranking | 1 search + 1 rerank pass | Does top-k match what answers actually use? |
| Guardrail checks | 1-3 evaluations | Which checks are needed at which stage? |
| Generation with context | 1 call with retrieved tokens | How much retrieved context is read vs wasted? |
| Agent tool follow-ups | 0-5 additional invocations | What is the loop stop condition? |
Illustrative pipeline composite. The audit measures your actual calls per answered query and prices each stage from your Bedrock usage data.
How It Works
How the audit works
Step 1
Share spend and workload shape
Submit monthly spend range, provider mix, token volume, and what your Bedrock workloads actually do: chat, RAG, agents, extraction, or batch jobs. Takes minutes, no production access needed.
Step 2
Map cost to workflows
We break the invoice into cost per completed user action: prompt and output tokens, retries, retrieval, tool calls, and orchestration overhead — the 72% of AI cost that hides outside the model call.
Step 3
Rank levers by friction and savings
Each leak gets a lever — routing, caching, prompt compression, retry control, quantization, or a private break-even case — ranked by expected savings against implementation effort.
Step 4
Get the written read, then decide
You receive the first cost-leak read in writing. If the economics justify deeper work, the next step is a scoped engagement; if not, you keep the findings.
Start with spend, provider, and workload shape.
The audit form routes teams below $20K/month toward self-serve estimators and routes qualified spend into follow-up.
$47K → $28K
Case study: a Llama 3 70B production workload moved from 4 GPUs to 2 with INT8 quantization, KV-cache pruning, and serving changes — a 42% monthly cost cut with 2.3x throughput.
Read the full auditFAQ
Common questions
Why are AWS Bedrock costs high?
AWS Bedrock costs rise when applications use larger models than needed, send long contexts, repeat guardrail checks, run multi-step agents, or combine RAG retrieval with expensive model calls.
How do you optimize Bedrock agent costs?
Bedrock agent costs can be optimized by limiting tool loops, routing simple tasks to cheaper models, caching stable context, shortening prompts, and measuring cost per completed user action.
Should Bedrock workloads move to self-hosted models?
Some Bedrock workloads should stay managed. Predictable high-volume private workloads may justify cloud GPU or on-prem serving after break-even analysis.
How much does AWS Bedrock cost per month?
Bedrock has no platform fee — the bill is the sum of per-token model charges across every call your workflows make. For RAG and agent systems that means one user query can trigger many billable steps: embedding, retrieval-augmented generation, guardrail checks, and tool-driven follow-up calls. Monthly cost tracks workflow design more than traffic.
Why are my Bedrock agent costs so high?
Agent frameworks bill every step: each tool call, each reasoning turn, each retry, and each guardrail evaluation is a separate model invocation. An agent that loops five times with two guardrail checks per turn can cost 10-15x a single completion. Stop conditions and loop budgets are the levers.
Is Bedrock cheaper than OpenAI?
It depends on the model mix and workflow, not the platform. Comparable-tier models price similarly across providers; the real difference comes from routing — matching each task class to the cheapest adequate model — and from orchestration overhead. The audit compares your cost per completed action across providers rather than list prices.