AWS Bedrock pricing: what each model costs today.

IHow does Amazon Bedrock pricing work?

You pay for what the model reads and what it writes. Tokens are the unit.

Every request has input tokens (prompt, instructions, history, retrieved documents, tool results) and output tokens (what the model writes). AWS quotes text models per 1 million tokens, with a higher rate for output. Reading a lot and writing a little, as in classifying or extracting, is cheap.

Thinking counts as output. AWS bills a Claude model's thinking tokens as output tokens, even when the response shows only a summary (AWS docs on extended thinking). The effort setting changes how much the newer Claude models think.

Third-party models such as Claude are billed through AWS Marketplace. They show on your AWS bill and in Cost Explorer under the model provider, not under Amazon Bedrock (Claude Haiku 5.5 model card).

The Bedrock pricing page lists no free allowance for model calls. New AWS accounts get $100 in Free Tier credits and can earn up to $100 more over 6 months (AWS Free Tier).

IIWhat does each model cost on Bedrock?

Standard rates per 1 million tokens, as AWS listed them on October 10, 2026 with the Region set to US East (Ohio). Prices change often and differ by Region.

Source: Amazon Bedrock pricing, checked October 10, 2026. Global cross-Region inference can send a request to any Region. Geo and In-Region inference keep it inside one geography or one Region, and on the newer Claude models they cost 10% more. Model details are on the AWS list of Anthropic models.

IIIWhich way of paying fits: Standard, Flex, Priority, Batch or Reserved?

The same model can cost half as much, or 75% more, depending on how you call it. Not every model supports every option.

You choose Standard, Priority or Flex per request (AWS service tiers). Reserved is set per account: Claude Sonnet 4.6 Reserved (Global, 1-month) was $0.18 an hour per 1K input tokens per minute and $0.90 for output. Batch is a separate job with results in Amazon S3.

IVHow much does prompt caching save on Bedrock?

Store a repeated prefix once, then reuse it for a fraction of the price.

Mark a prefix (a system prompt, tool definitions or a long document) as cacheable. On the Claude 5.5 models a cache checkpoint needs at least 512 tokens, with up to 4 per request (Claude Sonnet 5.5 model card; AWS prompt caching docs).

A 5-minute cache write costs 25% more than normal input, and a 1-hour write costs twice as much. Reads cost 10% of normal input or less. So a 5-minute cache pays off on the second request that reuses the prefix, and a 1-hour cache on the third. On Amazon Nova, AWS says cached reads cost 75% less than normal input.

VWhat else shows up on a Bedrock bill?

Tokens are usually the biggest line, not the only one. These prices are on the same AWS page.

• Embeddings: Billed per input token. Amazon Titan Text Embeddings V2 is $0.02 per 1M tokens and Cohere Embed 4 is $0.12.
• Images and video: Amazon Nova Canvas is $0.04 per standard 1024 x 1024 image. Amazon Nova Reel is $0.08 per second of 720p video.
• Knowledge Bases: A managed knowledge base costs $5.00 per GB of raw data per month plus $1.00 per 1,000 Retrieve calls. Parsing, embeddings and re-ranking with the managed models are included.
• Guardrails: Content filters cost $0.15 per 1,000 text units of up to 1,000 characters each. Sensitive information filters cost $0.10.
• Custom models: Fine-tuning adds training, monthly storage and hourly hosting. Hosting a custom Amazon Nova Canvas model costs $60.50 an hour per model unit with no commitment, whether or not you send requests.

VIWhat does a real workload cost per month?

Estimates from the rates above, at Standard Global prices. Your token counts will differ.

Support chatbot on Claude Haiku 5.5. 50,000 chats a month at 2,000 input and 400 output tokens each: about $20.00. Caching a 1,500-token system prompt cuts the input part from $10.00 to about $3.25, plus the cache writes.

Document assistant on Claude Sonnet 5.5. 20,000 questions a month at 6,000 input tokens (mostly retrieved passages) and 700 output tokens: about $380, plus about $30.00 for a managed knowledge base with 2 GB of data. The same traffic is about $19.00 on Haiku 5.5 and $760 on Opus 5.5.

Nightly tagging on Amazon Nova Micro. One million documents a month at 1,000 input and 50 output tokens: about $42.00 at Standard rates, or $21.00 as a Batch job.

Agent on Claude Opus 5.5. 2,000 tasks a month, each reading 30,000 tokens and writing 5,000 with its thinking: about $440. A lower effort setting means fewer thinking tokens.

VIIHow can you cut a Bedrock bill yourself?

Most of this needs no partner and no program.

• Pick the cheapest model that passes your test: Run 50 to 100 real prompts on Haiku 5.5 or Nova first and move up only the requests that fail. Switching models is a model ID change.
• Cache what repeats, batch what can wait: System prompts and tool definitions go in the cache. Offline jobs go to Batch or Flex, if your model supports them.
• Use the Global endpoint unless data must stay put: It is 10% cheaper on the newer Claude models.
• Cap output, trim context, lower effort: Every input token is billed on every request, and thinking is billed as output.
• Watch the bill: Estimate with the AWS Pricing Calculator, set an AWS Budgets alert and check Cost Explorer each week.

You can also apply for credits yourself at aws.amazon.com/startups/credits. AWS Activate Founders gives $1,000 to start, up to $5,000. AWS Activate Portfolio goes up to $200,000 with the Org ID from an Activate Provider that backs you. AWS says Activate applicants hear back within 5 to 10 business days.

VIIIWho pays for Bedrock if you are a startup, and who files it?

AWS credits can pay a Bedrock bill. The question is which program fits you and who applies.

Activate credits can pay for third-party models on Bedrock, Claude included (AWS announcement, April 2024; Activate terms). You apply for Activate yourself. A consulting partner cannot file it for you.

For a bigger build, a vetted AWS partner can file funding only partners can request: POC funding to test a Bedrock workload, IW funding when you already run on AWS and launch an AI feature, or MAP funding when you move to AWS. For startups it is usually $50K to $150K. Depends on the workload; AWS approves. That is far more than AWS Activate Founders gives.

CloudRoute matches you, for free, with one vetted AWS partner who can file it. We reply within 24 hours; your matched partner reaches out within 5 business days.