# AWS Bedrock pricing in 2026: What it really costs and how to optimize it ![Sinjan Ballav](https://cdn.manageengine.com/sites/meweb/images/cloudspend/sinjan-ballav.png) **Sinjan Ballav** FinOps Advocate and Cloud Economics Writer Last updated: August 2026 15 min read The AWS Bedrock pricing page looks straightforward. You see a list of models, their per-token rates, and a few billing modes. Most teams do the math, set a budget, and move forward with confidence. However, the first bill tends to change that. For most teams, actual costs run one-and-a-half to two times their initial estimates. This is not because AWS hides fees. The pricing is all there, technically speaking. The problem is that the real cost of running Bedrock in production is spread across four or five different services, amplified by agent behavior that is easy to underestimate, and shaped by a handful of default settings that most teams never revisit. This guide walks through token rates, billing modes, add-on services, adjacent AWS charges, and the optimization decisions that have the most impact. Use it to understand what Bedrock costs, why it costs that much, and where to look when you need to bring the cost down. ## What is Amazon Bedrock? Amazon Bedrock is a fully managed service from AWS that provides access to foundation models from Anthropic, Meta, Mistral, Cohere, Amazon, and others through a single API. The service is serverless; integrates with IAM, CloudWatch, and S3; and keeps your data within your AWS account boundary. Beyond raw model access, Bedrock includes several platform features that affect both cost and capability: - **Knowledge Bases** for retrieval-augmented generation (RAG), connecting models to your private data - **Agents and AgentCore** for autonomous, multi-step workflows where the model can plan and use tools - **Guardrails** for content filtering, PII detection, and safety policies applied across all models - **Flows** for visual orchestration of multi-step AI pipelines - **Model Evaluation** for programmatic and human-reviewed quality testing - **Prompt caching and Intelligent Prompt Routing** for built-in cost and latency optimization Common use cases include conversational assistants, RAG-based research tools, document processing pipelines, autonomous agents, code generation, and high-volume content generation. Each workload has different cost characteristics, which is where pricing complexity starts. ## How Amazon Bedrock pricing works A few variables drive your Bedrock bill, and understanding them is the foundation for everything that follows. **Token volume:** Bedrock bills per token for text models, where one token is roughly four characters of English text. Input tokens (your prompt) and output tokens (the model's response) are priced separately, with output typically costing three to eight times more than input. **Model choice:** Each model family has its own pricing, and the spread across the catalog is wide. The cheapest text model on Bedrock costs over 100 times less per token than the most capable one. **Operation type:** Image generation charges per image rather than per token. Embedding models charge on input tokens only. Video generation has its own pricing structure entirely. **Billing mode:** On-demand, batch, provisioned throughput, and prompt caching represent meaningfully different price points for the same underlying model. **Add-on services and adjacent AWS charges:** Knowledge Bases, Agents, Guardrails, and Flows each carry their own charges. CloudWatch, S3, Lambda, and data transfer add further costs that do not appear on the Bedrock pricing page. ## The mental model: Four layers of cost Before getting into specific numbers, it helps to understand how Bedrock billing is structured. Most cost overruns happen because teams focus on only one layer while the others quietly accumulate. **Layer 1: Inference.** The per-token charge for calling a model. This is what the pricing page leads with. **Layer 2: Billing mode.** Whether you use on-demand, batch, provisioned throughput, or prompt caching has a significant effect on the effective rate. The difference can reach 90% on certain workloads. **Layer 3: Add-on services.** Knowledge Bases, Agents, AgentCore, Guardrails, and Flows each carry their own charges on top of inference. These are the costs that most surprise teams in their first few months. **Layer 4: Adjacent AWS services.** CloudWatch, S3, Lambda, OpenSearch, and data transfer charges do not appear on the Bedrock pricing page, but they do appear on the AWS bill. Getting all four layers right removes most of the surprises. Missing any one of them is usually what turns a manageable bill into an uncomfortable conversation. ## Understanding the model landscape The Bedrock model catalog in 2026 spans roughly two orders of magnitude in cost between the cheapest and most expensive options. The same workload can cost 10 times more than it needs to, depending entirely on which model the team reached for first. Rather than reproducing AWS' pricing table (rates change every few months as new versions launch and older ones reprice; see the [official AWS Bedrock pricing page](https://aws.amazon.com/bedrock/pricing/) for current numbers), it is more useful to understand how the model families compare to each other and where each one fits. ### The cost tiers Bedrock models fall into roughly four cost tiers. Understanding which tier a model belongs to is more durable knowledge than memorizing specific rates. - **Premium tier:** Claude Opus and the larger Mistral models sit here. These are priced for complex reasoning, long-context analysis, and tasks where the cost of a wrong answer exceeds the cost of computation. - **Balanced tier:** Claude Sonnet, DeepSeek v3, Cohere Command R+, and Amazon Nova Pro offer strong general capability at a fraction of the premium tier cost. For most production chatbots, RAG applications, and general assistants, this is the right starting point. - **Efficient tier:** Claude Haiku, Nova Lite, and Mistral Small handle a surprising amount of real-world workloads, including classification, summarization, and structured extraction, at a meaningful discount to the balanced tier. - **Ultra-low-cost tier:** Nova Micro and Gemma 3 are priced at near-commodity rates. They are not suited for complex reasoning, but for the workloads they handle well, batching huge volumes becomes economically viable. ### Representative rates To give a sense of scale across the tiers, here are some illustrative examples from the current catalog. *Note that software and pricing information was last verified in August 2026. Features, pricing, and availability may have changed. Please verify current details with AWS before making decisions.* | Model | Input (per 1M tokens) | Output (per 1M tokens) | Tier | |---|---|---|---| | Amazon Nova Micro | $0.035 | $0.14 | Ultra-low-cost | | Google Gemma 3 4B | $0.04 | $0.08 | Ultra-low-cost | | Amazon Nova Lite | $0.06 | $0.24 | Efficient | | Claude Haiku 4.5 | $1.00 | $5.00 | Efficient | | Mistral Small | $0.20 | $0.60 | Efficient | | DeepSeek v3 | $0.27 | $1.10 | Balanced | | Claude Sonnet 4.5 | $3.00 | $15.00 | Balanced | | Amazon Nova Premier | $2.50 | $12.50 | Balanced/Premium | | Claude Opus 4.6 | $5.00 | $25.00 | Premium | The spread tells the story clearly. A team processing 100 million input tokens per month pays roughly $3.50 using Nova Micro. The same volume routed through Claude Opus costs $500. Both can be the right choice for different workloads, but the decision should be deliberate rather than incidental. ### A few patterns worth knowing **Claude pricing on Bedrock matches Anthropic's direct API:** AWS does not add a markup for Anthropic models routed through Bedrock. The practical trade-off is that new Claude versions typically arrive on Bedrock one to four weeks after they are available directly through Anthropic. **Open-source models carry a premium on Bedrock:** Llama and similar open-source models cost two to three times more on Bedrock than on dedicated inference providers like Together AI or Fireworks. The premium pays for IAM integration, VPC isolation, and HIPAA and FedRAMP compliance coverage. For regulated industries, that package often justifies the cost. For workloads without compliance requirements, the premium is harder to justify. **Amazon Nova is priced aggressively:** AWS positions the Nova family as the most cost-effective option in each tier. For workloads where Nova's capabilities are sufficient, it is often the cheapest path. The honest approach is to test whether it meets your quality bar rather than assuming either way. ## The four billing modes Most teams default to on-demand and never look back. That works for early-stage workloads, but it becomes an expensive habit at scale. Understanding the four billing modes and when each applies is where most teams find their first meaningful savings. ### On-demand You pay per token with no commitment or minimum. On-demand is the right choice for unpredictable workloads, multi-model experimentation, and anything where usage volume is not yet established. The trade-off is that it carries the highest per-token rate of any billing mode. If you are spending more than $40 per day on a single model consistently, you have likely outgrown it. AWS also offers [Flex and Priority tiers](https://aws.amazon.com/bedrock/service-tiers/) within on-demand. Flex provides cost savings for workloads that can tolerate some variability, while Priority charges a premium in exchange for faster performance and prioritized compute. ### Batch Batch processes requests asynchronously and delivers results within 24 hours at 50% of on-demand rates, with no change to output quality. It is the right choice for document processing pipelines, nightly summarization, bulk classification, and evaluation runs. The only catch is that it requires designing for async, so applications built around synchronous responses will need some rework. Example: Ten thousand nightly contract summaries costs roughly $500 per month using on-demand. On batch, it's $250 per month. That's $3,000 per year saved with no quality trade-off. ### Provisioned throughput Provisioned throughput lets teams purchase dedicated model capacity in advance, with one-month or six-month commitments and discounts of 15%–40% off on-demand rates. Specific rates are not published publicly, so contact your AWS account team for details. The break-even rule of thumb: If daily spend on a single model exceeds $30–$40 on on-demand, the math is worth running. Provisioned is per-model, so it works best for steady, predictable workloads rather than dynamic multi-model routing. ### Prompt caching When an application repeatedly sends the same content across requests, such as a long system prompt, a static RAG chunk, or a reference document, prompt caching stores those tokens so they do not need to be reprocessed on each call. Cache reads cost up to 90% less than standard input rates, making it the highest-ROI optimization available in Bedrock for most production workloads. The one structural requirement is that cacheable content must appear at the beginning of the prompt, before any dynamic user input. Example: A 4,000-token system prompt across 50,000 queries per day costs roughly $600 per month without caching. With an 85% cache hit rate, that drops to under $100 per month. ## Add-on services: Where the real surprises live This is where most budget overruns originate. Each add-on service looks like a feature addition on paper, but in practice each one stacks costs on top of inference. Knowing what each one charges, and what triggers those charges, is the difference between predictable bills and unpleasant ones. ### Knowledge Bases (RAG) Bedrock Knowledge Bases connect foundation models to private data through RAG. Inference tokens are billed at standard model rates, but the vector store sitting underneath is a separate cost with its own minimum. The most common surprise is the OpenSearch Serverless floor. When teams use the default vector store, the minimum monthly cost runs roughly $345–$350 regardless of how much data is stored or queried, because AWS requires a minimum of two OpenSearch Compute Units (OCUs) at approximately $0.24 per OCU per hour to operate. A Knowledge Base spun up purely for testing, with only a few hundred queries run against it, still generates that line item. Full details are on the [Amazon OpenSearch Service pricing page](https://aws.amazon.com/opensearch-service/pricing/). The better alternative for most workloads is [Amazon S3 Vectors](https://aws.amazon.com/s3/features/vectors/), which is a newer vector storage option from AWS. S3 Vectors costs up to 90% less than OpenSearch Serverless at trillion-vector scale, and for any new Knowledge Base project in 2026 it should be the default choice. For teams already running on OpenSearch Serverless, migration to S3 Vectors typically pays back within the first month. ### Bedrock Agents and AgentCore Agents are one of Bedrock's most capable features and one of its most expensive when not modeled carefully. When a user query reaches a Bedrock Agent, the model is called multiple times internally: once to plan, once per tool call, and once to synthesize the response. A query that triggers three tool calls can consume five to 10 times the tokens the prompt and response alone would suggest. Example: "Top five customers by revenue last quarter?" triggers planning (~500 tokens), a CRM query (~800 tokens), an analytics query (~900 tokens), and synthesis (~600 tokens). That is 2,800 tokens for what looks like a 200-token question. AgentCore uses consumption-based billing measured in per-second vCPU and gigabyte hours plus Gateway operation fees. See the [AgentCore pricing page](https://aws.amazon.com/bedrock/agentcore/pricing/). Budget against agent runtime duration, not token volume alone. ### Guardrails Guardrails are billed at $0.15 per 1,000 text units (one text unit equals 1,000 characters), an 80% reduction from the original 2024 rate. At scale, charges still add up. A chatbot handling 100,000 conversations per day at 2,000 characters per request generates $900 per month in Guardrails charges alone. Applying Guardrails to both input and output doubles the per-request charge, so many teams limit filtering to inputs only or to customer-facing endpoints. ### Bedrock Flows Flows is billed at $0.035 per 1,000 node transitions, metered daily. Charges are negligible at low volumes but accumulate quickly with complex graphs. Keep Flows graphs as flat as possible by combining consecutive nodes into single model calls where you can. ### Model evaluation LLM-as-judge evaluations bill at standard inference rates. Human evaluation tasks cost $0.21 each. Most teams use LLM-as-judge for volume and reserve human tasks for targeted spot-checks. ## The AWS bill nobody warned you about Bedrock costs are distributed across several AWS services that are not specific to AI, and these charges do not appear on the Bedrock pricing page. They are predictable once you know to look for them, and they belong in any honest cost model. ### CloudWatch logging Bedrock integrates with CloudWatch for prompt and response logging, billed at $0.50 per gigabyte ingested. Verbose system prompts logged on every request are the most common cause of unexpectedly high CloudWatch bills. At 100,000 requests per day, even moderate prompt lengths produce gigabytes of log data each month. The practical fix is to log selectively: errors and flagged responses by default, routine traffic sampled at 5%–10%. For compliance-mandated full logging, routing to S3 is roughly 10 times cheaper than CloudWatch ingestion. ### Amazon S3 Batch inference I/O, custom model artifacts, and Knowledge Base source documents all accumulate S3 costs that do not appear on the Bedrock pricing page. None are large individually, but together they add up. ### Lambda Agents that use Lambda functions as action group tools generate Lambda invocations on every tool call. The free tier covers one million requests per month, but high-volume agent deployments exceed it quickly. Full pricing is on the [AWS Lambda pricing page](https://aws.amazon.com/lambda/pricing/). ### Cross-region inference This adds a 10% surcharge to all token pricing. For most workloads in 2026, the primary region has sufficient capacity and the surcharge is pure overhead. A team running 50 million tokens per month pays an extra $540 per year for a feature they enabled during testing and never turned off. ### VPC and PrivateLink Enterprise deployments via [AWS PrivateLink](https://aws.amazon.com/privatelink/pricing/) typically cost $50–$150 per month per endpoint, depending on traffic volume. Getting visibility across all four layers is where many teams struggle. The charges spread across different AWS service categories and do not aggregate automatically. ManageEngine CloudSpend pulls together spend across all relevant services; filters by team, application, or environment; and surfaces anomalies before they compound into a large bill. ## Bedrock vs. direct API: The honest comparison For Claude models, the cost is identical either way. AWS does not add a markup for Anthropic models on Bedrock. The only trade-off is that new Claude versions typically arrive one to four weeks later than on Anthropic's direct API. For open-source models, Bedrock is two to three times more expensive than dedicated inference providers like Together AI or Fireworks. That premium covers unified billing, IAM integration, HIPAA and FedRAMP compliance, and no infrastructure to manage. For regulated industries, the compliance coverage alone often justifies the cost. For workloads without those requirements, going direct saves meaningfully. For most enterprises already running on AWS, the consolidation value is real. A hybrid approach—tier-one models through Bedrock for compliance and open-source models through dedicated providers for bulk processing—often produces the best overall economics. ## Real-world cost scenarios on Bedrock Applying current rates to representative workloads makes the cost structure more concrete. The scenarios below use realistic stack configurations and include the adjacent AWS service costs that most estimates omit. *Note: Dollar figures below are illustrative, based on Bedrock rates as of August 2026. The relative cost distribution across components, which is the more useful takeaway, holds regardless of how the underlying rates shift.* ### Scenario 1: Simple inference workload **Stack:** An efficient-tier model using on-demand billing, no Agents, and no Knowledge Base. A straightforward conversational assistant with no retrieval or orchestration layer. | Component | Monthly Cost | |---|---| | Inference (20M input + 8M output tokens) | $60 | | CloudWatch logging | $1 | | S3 (minimal) | $2 | | **Total** | **~$63/month** | This is a genuinely low-cost starting point. The risk is not the current stack but what happens when Agents or a Knowledge Base get added without first modeling the cost impact. Adding an OpenSearch-Serverless-backed Knowledge Base to this configuration would roughly triple the monthly cost immediately. ### Scenario 2: RAG application with Knowledge Base **Stack:** A balanced-tier model using on-demand billing, a Knowledge Base on S3 Vectors, and Guardrails on inputs only. A production assistant grounded in private data. | Component | Monthly Cost | |---|---| | Inference (50M input + 20M output tokens) | $450 | | S3 Vectors storage (10GB) | ~$30 | | Guardrails (inputs only, 100M text units) | $15 | | CloudWatch (sampled) + S3 documents | $4 | | **Total** | **~$500/month** | This is what a well-optimized RAG application looks like. The team chose S3 Vectors over OpenSearch Serverless, saving more than $300 per month on the vector store alone. Guardrails are applied to inputs only, and CloudWatch logging is sampled rather than complete. The same stack running on OpenSearch Serverless with Guardrails on both sides would cost $900–$1,000 per month. At enterprise scale with Bedrock Agents, the math gets more interesting. Agent token amplification typically multiplies effective inference costs by five times or more, and high-volume agent platforms easily reach $2,500–$3,000 per month before optimization. The same workload optimized across all four layers often runs at 60% of that cost or less. ## Six cost optimization strategies to implement These are not edge-case optimizations. They are the changes that consistently produce the largest reductions in real Bedrock deployments. ### 1. Use the cheapest model that passes your quality bar Most teams pick a model during development and never revisit the decision. A more structured approach is to run a model ladder test for each significant workload, starting at the ultra-low-cost tier and moving up only when quality falls below the threshold. AWS's Intelligent Prompt Routing can automate this within a model family, routing simpler queries to cheaper models automatically. ### 2. Enable prompt caching for repeated context If your application sends the same system prompt or context chunks on every request, prompt caching delivers up to 90% cost reduction on those cached tokens. The one structural requirement is that cacheable content must appear at the beginning of each prompt, before any dynamic user input. ### 3. Migrate Knowledge Bases to S3 Vectors Teams running a Knowledge Base on OpenSearch Serverless and paying the $350 per month floor have a clear path to savings. Migrating to S3 Vectors typically takes a few hours and reduces vector storage costs by up to 90%, with the savings starting immediately. ### 4. Move batch-compatible workloads off on-demand Document processing pipelines, nightly summarization, bulk classification, and evaluation runs do not need synchronous responses. Moving these workloads to batch cuts the cost in half with no change to output quality. ### 5. Audit and disable cross-region inference Cross-region inference adds a 10% surcharge to all token pricing, but most workloads in 2026 do not need it. If your primary region has sufficient capacity, disabling cross-region routing removes the surcharge without affecting availability. ### 6. Right-size agent orchestration For teams using Bedrock Agents, logging actual token consumption per agent invocation rather than per user query gives a clearer picture of true costs. If the average amplification factor is above five, there is room to reduce it through flatter agent graphs, batched tool calls, more concise system prompts, and routing simple queries away from agent orchestration entirely. ## Pre-production cost checklist Running through these questions before a new Bedrock workload goes to production catches most of the issues that lead to unexpected bills. **Model selection** - Have we tested the cheapest tier that meets our quality bar? - Are we constraining output length where possible? **Billing mode** - Does this workload need synchronous responses, or could it run as batch? - Are we above the provisioned throughput break-even threshold? - Is prompt caching enabled for repeated content? **Add-ons** - If we need a Knowledge Base, are we using S3 Vectors over OpenSearch Serverless? - Have we modeled the agent token amplification factor? - Are we applying Guardrails to both sides when one side would suffice? **Adjacent services** - Is CloudWatch logging sampled rather than complete? - Is cross-region inference disabled unless we need it? - Have we budgeted for Lambda, S3, and PrivateLink alongside Bedrock inference? **Ongoing** - Are AWS Budget alerts configured for Bedrock spend? - Are resources tagged for cost attribution by team, environment, and model? - Is Bedrock spend visible alongside the rest of our AWS costs in a single dashboard? ## Key takeaways AWS Bedrock pricing is manageable once all four layers are visible. The surprise costs are not hidden, just easy to miss: agent token amplification, the OpenSearch Serverless minimum, cross-region inference left enabled, and CloudWatch logging set to capture everything. Teams that keep costs under control choose models deliberately, model agent token costs before deployment, and treat the adjacent AWS bill as part of the Bedrock budget from day one. The savings are material. A mid-market RAG app at $1,000 per month often drops to $500 with a few targeted changes, and enterprise agent platforms frequently have 25%–30% in addressable waste. ## Manage your Bedrock costs with ManageEngine CloudSpend Understanding what Bedrock costs is one part of the problem. Having ongoing visibility into those costs across services, teams, and environments is the other. Most teams start by piecing together their AWS bill manually, correlating charges from Bedrock, CloudWatch, S3, Lambda, and OpenSearch across different sections of the AWS console. That works at small scale but stops being practical as usage grows. [ManageEngine CloudSpend](https://www.manageengine.com/cloudspend/) gives engineering, DevOps, and finance teams a unified view of cloud spend. You can filter Bedrock costs by team, application, environment, or cost center, and see the full picture across every related service in one place. CloudSpend also surfaces anomalies and cost trends proactively, so a spike in agent token consumption or an unexpected OpenSearch charge gets flagged before it turns into a budget conversation. If you are running Bedrock in production and want cost visibility that covers all four layers described in this guide, CloudSpend is built for exactly that use case. *Note: Pricing data verified against AWS official pricing pages as of August 2026. Model rates and service features change regularly. Confirm current numbers before finalizing budgets.* ## Frequently asked questions ### Is Amazon Bedrock pricing the same as calling the Anthropic API directly? For Claude models, yes. AWS does not add a markup for Anthropic models on Bedrock. Per-token rates match whether you call through Bedrock or directly through the Anthropic API. For open-source models like Llama, Bedrock is two to three times more expensive than dedicated inference providers. ### Why is my Amazon Bedrock bill higher than expected? The most common causes are agent token amplification (a single user query can consume five to 10 times the tokens you would expect), the OpenSearch Serverless minimum of roughly $350 per month on Knowledge Base deployments, cross-region inference left enabled after testing, and CloudWatch logging set to capture everything. The real cost of Bedrock in production is spread across four layers, only one of which appears on the Bedrock pricing page. ### What is the cheapest model on Amazon Bedrock in 2026? Amazon Nova Micro is currently the lowest-cost text model on the platform, priced at $0.035 per million input tokens and $0.14 per million output tokens. Google Gemma 3 4B is comparable. Both sit in the ultra-low-cost tier and are best suited for bulk structured tasks, classification, and extraction rather than complex reasoning. ### When should I use batch inference vs. on-demand billing on Bedrock? Use batch for any workload where a user is not waiting for a real-time response. Document processing pipelines, nightly summarization jobs, bulk classification, and model evaluation runs are all good candidates. Batch costs 50% of on-demand rates with no change to output quality. If your workload requires synchronous responses, stay using on-demand. ### How do I reduce Amazon Bedrock Knowledge Base costs? The single largest saving is accomplished by migrating from the default OpenSearch Serverless vector store to Amazon S3 Vectors, which costs up to 90% less. OpenSearch Serverless carries a minimum monthly cost of roughly $345 to $350 regardless of usage. S3 Vectors has no such floor. Migration typically takes a few hours and pays back within the first month. ### What is Amazon Bedrock prompt caching, and how does it save money? Prompt caching stores repeated content (like long system prompts, static RAG chunks, and reference documents) so it does not need to be reprocessed on every request. Cache reads cost up to 90% less than standard input rates. To activate caching, cacheable content must appear at the beginning of the prompt, before any dynamic user input. ### Does Amazon Bedrock add a markup over open-source model providers? Yes, for open-source models. Llama and similar models cost two to three times more on Bedrock than on dedicated inference providers like Together AI or Fireworks. The premium covers unified billing, IAM integration, HIPAA and FedRAMP compliance, and managed infrastructure. For regulated industries, this is often worth it. For workloads without compliance requirements, going direct saves meaningfully. ### How do I track and attribute Amazon Bedrock costs across teams? Tag all Bedrock-related resources by team, application, and environment, then use those tags to filter costs in your AWS Cost Explorer or a dedicated cloud cost management tool. The challenge is that Bedrock costs spread across multiple AWS service categories (Bedrock, CloudWatch, S3, Lambda, OpenSearch) and do not aggregate automatically. [ManageEngine CloudSpend](https://www.manageengine.com/cloudspend/) consolidates spend across all four layers into a single view, filtered by whatever dimensions matter to your organization. --- ### Author Spotlight ![Sinjan Ballav](https://cdn.manageengine.com/sites/meweb/images/cloudspend/sinjan-ballav.png) **Sinjan Ballav** FinOps Advocate and Cloud Economics Writer Sinjan covers cloud cost management and FinOps at ManageEngine. He writes about the financial realities of running cloud infrastructure, from optimizing cloud spend to fostering cost accountability across teams, helping organizations move from cloud cost chaos to confident, data-driven decisions.