Amazon Bedrock has no single monthly subscription price. A useful estimate combines the exact model and inference tier, input and output tokens, and any retrieval, storage or operational services you enable. Start with the selected model's regional rate card. Then calculate the application around it. A direct model call and a managed retrieval application can have very different bills.
For Claude pricing, distinguish the model version and global or regional endpoint. For Knowledge Bases, identify the actual vector store before assuming a monthly minimum. The frequently repeated “$700 minimum” comes from one AWS solution example, not a universal charge for every Bedrock application.
Our data: reproducible Bedrock cost scenarios
We calculate three token-volume scenarios, a selected-model cost example and a batch sensitivity example. We also separate the components of AWS's published $700.20 retrieval example. Inputs and formulas are shown so you can substitute your account's rates. This is documentation-based analysis checked October 28, 2026, with AI-assisted drafting; no PE Collective production benchmark is claimed.
Model rates · Token arithmetic · Batch and capacity · Knowledge Bases · Bedrock versus direct APIs · Budget checklist
Which Bedrock pricing table should you use?
The AWS Bedrock pricing page separates providers and models, and describes Standard, Flex, Priority and Reserved tiers. It also lists model-dependent batch pricing. Select the region and deployment route that match your application. Copying the cheapest visible rate into a spreadsheet without those labels produces an estimate you cannot reproduce.
| Record with the price | Why it changes the estimate |
|---|---|
| Exact model ID and version | A model family name can hide different generations and prices. |
| Region and endpoint scope | Global and regional routing can have different commercial terms. |
| Processing tier | Interactive, asynchronous and reserved capacity are different products. |
| Input, output and cache rates | Each billed category needs its own token count. |
| Modality and optional tools | Images, audio, search and other features can add separate units. |
| Observation date and account terms | Published list prices and negotiated rates can differ. |
Anthropic Claude on Bedrock: avoid the old blanket claims
Anthropic's pricing documentation lists direct API base rates of $3 input and $15 output per million tokens for Claude Sonnet 4.6, and $5/$25 for Opus 4.6. Those are reference points for comparison, not a quote for every AWS region or endpoint. The same documentation directs Bedrock buyers to AWS's rate card.
It also distinguishes global and regional partner endpoints and describes a regional premium for the applicable newer Claude models. The old claim that “cross-region always adds ten percent” reverses an important distinction: routing scope and the named endpoint determine which rate applies. Preserve the selected endpoint beside the price rather than applying one multiplier to every model.
If the selected AWS table does not expose the rate in your export, leave that cell unquoted until you can confirm it in the rate card or account estimate. A direct Anthropic rate is not evidence of an AWS invoice amount. This matters especially for older models, special processing modes and private offers.
A currently published alternative example
AWS's visible table lists MiniMax M2.5 at $0.30 input and $1.20 output per million tokens for US East (N. Virginia), US East (Ohio) and US West (Oregon). That provides a concrete example for the arithmetic below. It is a cost illustration, not a recommendation that this model can replace Claude for your workload.
Evaluate candidate models against the same acceptance criteria. If one model needs several retries or a human to repair its output, a lower token rate may not produce a lower cost per completed task. Keep model quality and price as separate measurements until you can combine them using observed completion rates.
Calculate input and output tokens separately
For ordinary text inference, the basic formula is input tokens / 1,000,000 × input rate, plus output tokens / 1,000,000 × output rate. Add other billed categories separately. “One million tokens” is a pricing unit; it is not one million requests or one million words.
Suppose a request sends 2,000 input tokens and generates 500 output tokens. At the MiniMax M2.5 regional rates above, the token charge is 2,000 / 1,000,000 × $0.30 + 500 / 1,000,000 × $1.20 = $0.0012. The input and output portions each contribute $0.0006 in this particular example.
| Monthly requests | Input tokens | Output tokens | Illustrative inference cost |
|---|---|---|---|
| 10,000 | 20 million | 5 million | 20 × $0.30 + 5 × $1.20 = $12 |
| 100,000 | 200 million | 50 million | 200 × $0.30 + 50 × $1.20 = $120 |
| 1,000,000 | 2 billion | 500 million | 2,000 × $0.30 + 500 × $1.20 = $1,200 |
These examples assume identical requests, no retries, no caching and no other services. Real traffic has a distribution. Measure typical requests and the long tail separately: a small number of very long conversations can account for a meaningful fraction of a monthly bill.
For a conversation, include the history resent on later turns. For retrieval, include the passages inserted into the prompt. For a tool-using workflow, include subsequent model calls that consume tool results. A user's single click can initiate several billable requests, so a request-level unit cost may understate the cost of a completed user task.
Use actual usage counters after a pilot. Tokenizers differ between models, and character-based estimates can behave differently for code, tables and non-English text. If you compare providers, tokenize the same representative inputs with the appropriate tokenizer or inspect the reported usage from each service.
Batch inference, service tiers and reserved capacity
AWS advertises a 50% reduction from on-demand inference for select models using batch inference. Confirm support for the specific model and region. The batch guide describes asynchronous jobs with input and output files in S3, so this mode fits work that can wait for job completion.
A document backfill, scheduled classification job or nightly summary queue may tolerate that workflow. A live chat response generally has a different latency requirement. Group work by deadline before applying a batch discount to the whole forecast; the discounted rate is irrelevant if it cannot meet the task's delivery time.
For sensitivity analysis, suppose an eligible workload would cost $200 in on-demand model charges. A 50% reduction makes the corresponding batch model charge $100: $200 × 0.5. If the application also has $60 of storage and other charges that do not change, the total moves from $260 to $160. That is a $100 saving, or 38.5% of the original total, rather than half the entire bill.
Batch also needs failure handling. Preserve a stable identifier for each input record, reconcile returned results and retry only failed work. Re-running a whole job because one record failed can erase some of the expected saving. Include S3 operations and retention in your architecture estimate.
Provisioned Throughput is a separate capacity decision. AWS's pricing page directs Anthropic customers to their account team for current provisioned pricing and shows model-specific arrangements elsewhere. Do not extrapolate an old hourly rate or commitment period across all providers. Ask for the exact capacity unit, term, supported model and utilization assumptions.
Compare reserved capacity with the workload's actual demand pattern. A system busy for a short daily interval can leave paid capacity idle for the rest of the day. A consistent workload may support a different choice. Calculate effective cost per accepted request under both expected utilization and a lower-demand scenario.
Bedrock Knowledge Bases pricing and the $700 example
A retrieval application can add document ingestion, embedding, vector storage, retrieval, reranking and answer-generation costs. Some components scale with traffic; others depend on provisioned capacity or retained data. The vector store you choose and its configuration are essential inputs to the budget.
The AWS Generative AI Application Builder cost example shows $691.20 per month for a particular OpenSearch Serverless configuration and $9 for query embeddings, totaling $700.20. It explicitly describes a sample configuration and notes that costs vary with the workload and existing resources.
That example does not establish a universal $700 Bedrock minimum. Reusing an existing store, choosing another supported configuration or running a simple model-only application changes the cost structure. Conversely, a larger corpus, higher availability requirement or more traffic can cost more than the example.
Separate one-time ingestion from recurring query work. A million documents embedded during migration create a different spending pattern from a small daily update. Estimate document changes, deletion handling and full reindexing events. If a new embedding model requires rebuilding the index, the migration month may carry both old and new storage costs.
Reranking is another independent line. Decide how many candidate passages are sent for reranking and whether the service's billing unit changes above a document-count or length boundary. Retrieving more candidates can help recall, but it can also increase latency and cost. Measure the additional useful answers before making that the default for every query.
A cost worksheet for a retrieval application
| Component | Measurement to collect | Budget treatment |
|---|---|---|
| Initial embeddings | Tokens after chunking and overlap | Migration or onboarding expense |
| Index updates | Changed tokens per day | Recurring ingestion |
| Vector store | Storage, capacity and replicas | Separate fixed and usage components |
| Query embeddings | Searches and tokens per search | Traffic-dependent model usage |
| Reranking | Requests and candidate passages | Optional retrieval stage |
| Generation | Prompt, retrieved context and output | Input/output token calculation |
| Operations | Logs, networking and application services | Service-specific charges |
Use a row only when the component exists in your design. Adding every AWS service to a prototype estimate can be as misleading as ignoring the operational layer entirely. The worksheet should mirror the architecture diagram and the eventual cost-allocation tags.
Is Bedrock cheaper than the direct model API?
Compare the same model version, geography, processing mode, input/output mixture and cache behavior. Then add the surrounding application costs on each side. There is no universal answer based only on the provider name.
For an AWS-native organization, central billing, permissions and existing operations can reduce integration work. For another team, a direct provider API may be easier to adopt and may expose a different set of features. Treat those as architecture considerations to verify against the product documentation, not assumptions that one route always has every feature first.
Check whether the model is generally available in the chosen region and whether the endpoint supports the required functionality. A price for a model your account cannot use does not belong in an executable budget. Likewise, a nominally cheaper route that lacks required output controls can force application changes whose cost outweighs the token saving.
Do not infer contractual data handling from the fact that a service appears on an AWS bill. Review the exact service, provider terms, region and deployment configuration. Marketplace-billed direct services and partner-operated Bedrock endpoints are distinct products; their commercial and technical rules can differ.
A regional cost difference you can reproduce
AWS's MiniMax M2.5 table gives a useful demonstration of why a regional label belongs beside every price. It lists $0.30 input and $1.20 output per million tokens for the three US regions named above, and $0.47 input and $1.86 output for Europe (London). These are documented prices for this specific model and route, not a universal regional multiplier for Bedrock.
Apply the same illustrative monthly workload of 200 million input tokens and 50 million output tokens. The US example is 200 × $0.30 + 50 × $1.20 = $120. The London example is 200 × $0.47 + 50 × $1.86 = $187. The difference is $67, or about 55.8% of the US inference charge. Storage, networking, taxes, retries and other services remain outside both calculations.
That calculation does not recommend moving workloads between regions. Data-location requirements, latency and model availability can determine the deployment before price is considered. Its purpose is to expose a spreadsheet error: using one region's unit rates while labelling the budget with another region. Repeat the check separately for every model you plan to call.
Build a Claude estimate without substituting the wrong rate
For the Claude queries this page serves, the required inputs are the exact AWS model identifier, endpoint scope, region, processing tier and separate input/output rates. Add cache-write and cache-read rates only if the selected model and request support them. Do not paste the direct Anthropic comparison rates into those AWS cells and silently treat them as verified Bedrock prices.
For example, a workload with 200 million ordinary input tokens and 50 million output tokens has the formula 200 × AWS input rate + 50 × AWS output rate. Once you have confirmed both rates in the regional AWS table, this formula produces the model-only estimate. If part of the input is billed as cached input, subtract those tokens from ordinary input and price them in the appropriate cache category instead; counting them in both rows overstates the bill.
Save the price-table selection with your estimate so another engineer can reproduce it. If a legacy Claude version is unavailable to the intended account, compare an available replacement as a separate scenario. A historical model price remains useful for explaining an old invoice, but it is not a deployable new-project budget.
Before approving a Bedrock estimate
- Save the selected model ID, region, inference profile and rate-card date.
- Measure input and output usage for representative requests, including retrieval context and retries.
- Separate interactive traffic from jobs eligible for asynchronous processing.
- List every storage and application service the design deploys.
- Calculate expected and high-usage months, with migration costs shown separately.
- Set spending alerts and assign an owner to investigate changes in cost per accepted task.
After a pilot, reconcile the forecast with actual billed categories. If the estimate is wrong, identify whether the cause is token volume, model selection, cache misses, additional calls or infrastructure. That explanation is more useful than changing the price table without understanding the workload.
For the next step, use the official AWS pricing page to fill the model-rate cells, then apply the formulas above. See our LLM API cost comparison for cross-provider arithmetic and embedding guide for retrieval-specific choices. Recheck volatile rates before purchase; this guide's editorial owner is PE Collective.
