AWS Bedrock pricing 2026: per-token cost and throughput.

01 Key findings

1. Bedrock is a marketplace, priced model-by-model. Each foundation model is priced separately by its provider and routed through the AWS billing line. The price spread is vast — Claude 3 Opus output tokens cost roughly 1,250x a Nova Micro input token.

2. Model selection is the single largest cost lever. A RAG pipeline defaulting to Opus where Haiku suffices inflates cost by up to 18x with no quality benefit on routine traffic. Cascade routing alone commonly delivers 55–70% savings.

3. Provisioned throughput only pays above break-even. For Claude 3.5 Sonnet, break-even sits around 8–10 million tokens per hour. Below it, on-demand is cheaper and simpler; above it, committed capacity wins.

4. The commitment trap is over-provisioning. A one-year model-unit commit captures 40–50% savings but bills whether used or not. Right-size to 70–80% of observed peak, not to headline demand.

5. List price is negotiable through the EDP, not the console. Bedrock consumption burns down EDP commit at full list; a Bedrock-specific discount sleeve can exceed the blended rate by 10–20% when Bedrock is the strategic spend.

02 Per-model token pricing

Each foundation model on Bedrock is priced separately by its provider. The rates below reflect AWS public pricing as of Q1 2026 for the us-east-1 region, per 1,000 tokens.

Claude 3 Opus at $0.075 per 1K output tokens costs 1,250 times more than Nova Micro at $0.00006 per 1K input tokens. That spread is why routing — not raw volume — determines most Bedrock bills.

03 On-demand vs provisioned throughput

Bedrock has two consumption modes. On-demand charges per token with no minimum commitment and shares regional model capacity across all AWS customers. Provisioned throughput reserves dedicated capacity, measured in model units (MU), at an hourly rate regardless of utilisation.

The break-even is workload-specific. For Claude 3.5 Sonnet, a single MU lists at $39.60 per hour — $28,512 per month on a one-month no-commit reservation. That MU produces roughly 16 million input plus 4 million output tokens per hour under typical mixed loads, so workloads consuming 8–10 million tokens per hour sit at the break-even point.

Provisioned throughput is the right answer for predictable, high-volume workloads where latency consistency matters. On-demand suits variable workloads, pilots, and anything below break-even. Mixing both is common: provisioned capacity for the steady-state floor, on-demand for spike absorption.

04 The commitment trap

The one-year commit is where over-buying does real damage. A committed MU bills 24x7 whether the traffic arrives or not, so a reservation sized to headline demand quietly becomes the most expensive line on the invoice during off-peak hours.

Custom Model Import raises the same trap in a different form. Running a fine-tuned open-weights model (Llama, Mistral, Flan-T5) on Bedrock is priced per-model-copy per-hour with a five-minute minimum plus storage. A custom Llama 3 70B copy lists at $33 per hour: a 24-hour, single-copy pilot costs $792, but a 2-copy, 24x7 production deployment lists at $48,180 per month before EDP discount, plus $1.95 per GB-month for the weights. Idle copies, like idle MUs, are pure waste.

05 Cost at scale

Model choice compounds with volume. A retrieval-augmented chatbot serving 50,000 queries per day at 4,000 input (knowledge context) and 600 output tokens per query shows the effect — identical workload, monthly list cost by routing strategy:

06 Agents, Knowledge Bases, Guardrails

The orchestration features carry no incremental feature fee — the cost is in the tokens and infrastructure underneath. Bedrock Agents bills the standard model rate multiplied by the number of model calls in the reasoning chain: a six-step agent incurs six times the token cost of a single-shot prompt. Knowledge Bases (RAG) is free as a feature; the cost is the vector store, embeddings and document storage.

Bedrock Guardrails (PII detection, prompt-injection detection, content filtering) is priced per text unit of 1,000 characters. A guardrail on a 2,000-character prompt and 1,500-character response costs $0.00075 per request — at 10 million requests per month, that adds $7,500 to the bill. Titan Embeddings and Cohere Embed both list at $0.0001 per 1K tokens; alternative vector stores (Pinecone, Redis) price separately.

07 Cost-control patterns

Five patterns reliably reduce Bedrock spend by 30–60% without compromising application quality.

1. Model cascade routing. Route the easy 80% of traffic to Haiku, Nova Lite or Llama 70B; reserve Sonnet, Opus or Llama 405B for queries that fail the lightweight model's confidence check. Cascade routing typically cuts token spend 55–70% on support and search use cases.

2. Prompt caching. Anthropic models on Bedrock cache at $0.00075 per 1K cached tokens read — 75% off standard input pricing. Long system prompts, knowledge-base context and few-shot examples cached across invocations slash the cost of high-volume agent loops.

3. Output token cap. Set max-tokens conservatively. The SDK default is often 4,096 output tokens; most production prompts produce 200–500 useful tokens. The cap prevents runaway generations from inflating cost up to 8x.

4. Batch inference. For non-real-time work (overnight summarisation, content review), Bedrock Batch lists at 50% off on-demand on supported models, running on an AWS-managed schedule with no completion-time SLA.

5. Reserved capacity with rightsizing. Above the on-demand break-even, one-year provisioned throughput captures 40–50% savings. Measure on-demand consumption for 30 days, then commit to 70–80% of observed peak, leaving on-demand to absorb the remainder. See our cloud cost optimization guide.

08 Cost-modelling framework

Model Bedrock cost along four axes before committing to any deployment architecture.

09 Bedrock vs SageMaker and direct API

Bedrock and SageMaker are not substitutes. Bedrock is the managed foundation-model service; SageMaker is the full ML platform — notebooks, training, model registry, inference endpoints and MLOps. Use Bedrock to invoke a foundation model via API for a generative application; use SageMaker to train, fine-tune, deploy or operate a custom model with full control of the inference stack. Many architectures use both.

SageMaker JumpStart sits between them, deploying open-weights models on dedicated endpoints priced per instance-hour rather than per-token. For very high-volume inference on a single open model, JumpStart often beats Bedrock economics; below 8–10 million tokens per hour, Bedrock is cheaper and operationally simpler. For the multi-vendor view comparing Bedrock against Azure OpenAI, Vertex AI and direct Anthropic, see our AI procurement guide, the AI contract clauses guide and the AWS vendor hub.

10 Recommendation

Bedrock is a strategic AWS line with explicit growth quotas, which gives buyers more room than they realise. Three moves work in 2026: a Bedrock discount sleeve separating it from general infrastructure spend; a price-lock protecting against upward repricing while preserving beneficial cuts; and negotiated data-residency and indemnity terms, which vary by model provider. See our AWS EDP negotiation playbook for the commit structure.