AWS Bedrock

Amazon Bedrock is AWS's managed service for calling foundation models from several labs through one API. You pay per million tokens, in one of four service tiers, and AWS says the model providers never see your prompts. On the AWS price list on 31 October 2026, Claude Haiku 4.5 on demand was $1.00 per million input tokens and $5.00 per million output. It is a fit for teams already on AWS.

Some links on this page go through /go/ and may earn us a fee; it never changes our order.

Facts at a glance

For other options, see the AI hosting overview and the interactive hosting compare.

What Bedrock is

Bedrock is a gateway to many models inside your AWS account. You enable the models you want in a region, call them with the AWS SDK, and get one invoice. The models come from several labs, and the FAQ lists Anthropic, Meta, Mistral AI and Amazon among them. It is not a GPU rental: you cannot log in to a machine, and you do not choose a card. If you want that, look at Runpod or Lambda Cloud.

Beyond raw inference, AWS bundles features around it:

• Guardrails. Configurable safeguards that the FAQ says apply across supported models, fine-tuned models and models hosted outside Bedrock.
• Knowledge bases. A managed route to retrieval over your documents, which handles ingestion and encryption.
• Customisation. Fine-tuning with tagged data or continued pre-training with untagged data, and Custom Model Import for select publicly available architectures. The FAQ lists Llama 2 and 3, Mistral and Flan architectures.
• Playground. A console chat for trying models.

The four service tiers

The service tiers page says Bedrock offers four tiers for model inference so you can optimise for availability, cost and performance.

You set the service_tier parameter to reserved, priority, default or flex on the runtime API call. There are also Batch inference, at 50 percent less than on-demand for select models, and Global Cross-region Inference, which routes across regions for capacity.

The prices

Bedrock's pricing page loads rates dynamically by model, region and tier, so a plain fetch shows only the headings. We read the published AWS price list for the US East (N. Virginia) region instead, on 31 October 2026. The list shows two on-demand rates for the same model, one about 10 percent above the other. We assume the higher rate is for in-region processing and the lower for global routing, but the list does not say so in the rows we read.

All in dollars per million tokens. Models from other labs, such as Llama, Mistral and Amazon's own, have their own rates on the pricing page.

Provisioned Throughput is priced per million tokens-per-minute hour, for example $0.054 an hour for 1M input TPM on Claude Haiku 4.5 on one of its rows. It suits steady, high-volume use where you want guaranteed capacity.

Cached input is cheaper: the pricing page says cache-read input tokens are 75 percent below the on-demand input price. A long, repeated system prompt then costs a quarter of its list rate.

What you pay in practice

The volumes are examples we chose: 40 million input and 8 million output tokens a month, like the other pages in this section.

A second example: the same Haiku workload with 80 percent of input served from the cache. Cached input is priced 75 percent below on-demand, so $0.25 per million. 32 million cached tokens cost $8.00, the 8 million uncached $8.00, and output $40.00, for $56.00 against $80.00.

Compare that with open models on a per-token API: gpt-oss-120b at $0.15 and $0.60 on Groq, or $0.35 and $0.75 on Cloudflare Workers AI, costs $10.80 and $20.00 for the same volume. Frontier models on Bedrock cost several times more. The question is whether the quality gap is worth it for your task, and that is a test you run on your own prompts.

Bedrock has no GPU hours to compare, but a rented GPU such as a Runpod Secure A100 at $1.79 an hour costs about $1,289 for a month. Bedrock's per-token model is cheaper for light use and loses at very heavy steady use, as with other APIs.

Data and security

The data protection page says Bedrock has a concept of a Model Deployment Account: in each region where it is available, there is one such account per model provider, owned and operated by the Bedrock service team. After a model provider delivers a model to AWS, Bedrock deep-copies the provider's inference and training software into those accounts. Because model providers do not have access to those accounts, they do not have access to Bedrock logs or to customer prompts and completions.

The FAQ adds that when you fine-tune a model, your data is never exposed to the public internet, never leaves the AWS network, is transferred through your VPC, and is encrypted in transit and at rest, with the same AWS access controls as other services. You can protect traffic with Amazon VPC and AWS PrivateLink.

The same page includes a general warning to never put confidential or sensitive information in free-form fields such as names.. The compliance standards Bedrock supports are listed on the FAQ.

For the strictest data rules, an AWS contract and your own review matter more than any summary here. See private AI for teams.

Regions

Bedrock is offered in specific regions, and each model is offered in a subset. The FAQ points to a regions list in the docs. Cross-region inference lets requests route to other regions for capacity, which affects where data is processed, so check that setting if residency matters.

Limits

• Model access. You enable each model per account and region, and some need an approval step.
• Quotas. Throughput quotas apply per model and region, with increases through support.
• Reserved tier minimums. The Reserved tier has minimum capacity requirements and goes through your AWS account team.
• Pricing complexity. Rates vary by model, region and tier, so budgeting needs the price list, not one number.
• Not a GPU rental. You cannot run an arbitrary model unless it is supported or imported.

Choosing a tier

The tier you pick changes both cost and behaviour, so match it to the traffic, not to habit.

A worked comparison with our example volumes: Claude Sonnet 4.6 on demand at 40 million input and 8 million output tokens costs $240. Moving the 70 percent of that work that can wait to batch brings that portion from $168 to $84, a saving of $84, so the month costs about $156. The change is a parameter and a queue, not a rewrite.

Building agents on Bedrock

Agents are loops of model calls and tool calls, and each loop iteration is billed in tokens. Three habits keep the bill predictable.

1. Cache the system prompt and tool definitions. They repeat on every call, and cache reads cost 75 percent less than normal input.
2. Put long-running agents on Flex or batch where latency does not matter.
3. Meter per task. Log input and output tokens for each agent run, so a stuck loop shows up as one expensive task and not a surprise at month end.

For the wider build pattern, see how to build an AI agent. If your agents call tools over MCP, our list of AWS MCP servers covers the AWS side.

Cost controls

• AWS Budgets and alarms on the Bedrock service line.
• Per-role limits. Give each application its own IAM role and tag usage, so the bill can be split.
• Model choice per step. Use the cheaper model for routing and classification, and the stronger one for the final answer. Haiku 4.5 at $1 and $5 per million is a third of Sonnet 4.6's rate on both input and output.
• Output caps. Output costs five times input on these rows, so set a maximum output length.
• Review the price list quarterly. Model rates and the model list change, and new models can undercut old ones.

Leaving Bedrock

Because Bedrock uses its own SDK calls, moving to another API takes more code change than moving between OpenAI-compatible hosts. Wrap your model calls in one function, keep prompts in files you own, and keep an evaluation set so that you can compare a new provider on your tasks. If you fine-tuned a model on Bedrock, check what you can export. . AWS gateways and open-source gateways that put several providers behind one API, as listed in OpenRouter alternatives, reduce that lock-in.

Who should pick Bedrock

Alternatives

• Hugging Face Inference: dedicated endpoints plus a provider router.
• Cloudflare Workers AI: cheap open-model tokens.
• Modal and Replicate: per-second billing.
• Runpod and Lambda Cloud: your own GPU.
• Gateways: OpenRouter alternatives.
• Direct vendor APIs: see Anthropic vs OpenAI API.

Getting started

1. In the Bedrock console, open model access and enable the models you want in your region.
2. Create an IAM role with least-privilege access to the runtime API.
3. Call the model with the AWS SDK, setting service_tier if you want something other than Standard.
4. Turn on Guardrails if the app faces users.
5. Use batch for offline jobs, which is 50 percent cheaper for select models.
6. Set budgets and alarms in AWS Cost Management, since token bills grow with usage.

Our take

Bedrock sells governance and convenience more than price. If your company already lives in AWS, the IAM, VPC and billing fit is worth real money, and the statement that model providers cannot see prompts is a strong one. If you do not, a direct API or an open-model host will usually cost less and be quicker to start. Use batch and caching to cut bills by half or more where your workload allows. Our conclusions rest on AWS documentation and its published price list. We did not call the service for this page.

Related pages

• Hosting hub and AI hosting overview
• Cloud vs self-hosted AI
• Private AI hub