Amazon Bedrock

About Amazon Bedrock

Amazon Bedrock is AWS's managed platform for building generative AI applications and agents with foundation models from multiple providers. It combines model access with evaluation, retrieval, guardrails, customization, data tooling and agent infrastructure while fitting into AWS identity, networking, logging and billing. This page focuses on Bedrock itself rather than the wider AWS cloud catalogue.

What does Amazon Bedrock do?

Amazon Bedrock gives developers API and console access to foundation models without having to operate the underlying model-serving infrastructure. Teams can select models for text, reasoning, coding, image, speech and other supported workloads, then invoke them from applications using AWS-managed endpoints. AWS currently supports models from Amazon and multiple external providers, with availability varying by Region and model.

Bedrock also provides tools around the model call. These include model evaluation, Knowledge Bases for retrieval-augmented generation, Guardrails, prompt management and optimization, model customization, data automation, batch processing and several service tiers for balancing latency and cost. This makes Bedrock more than a model reseller: the buying decision is about using AWS as the operating layer around generative AI.

How broad is the model choice in Bedrock?

The current Bedrock catalogue includes Amazon Nova and models from providers such as Anthropic, Cohere, DeepSeek, Google, Meta, Mistral AI, NVIDIA, OpenAI, Qwen and others. Bedrock Marketplace extends the catalogue with more than 100 additional popular and specialized models that can be discovered and deployed through managed endpoints.

Model choice is one of Bedrock's strongest advantages for teams that do not want one application permanently tied to one provider. It is not a guarantee of portability, however. Prompt behavior, tool use, context limits, multimodal support, pricing and regional availability differ among models, so switching models can still require testing and application changes.

How do agents work in Amazon Bedrock in 2026?

AWS's agent architecture changed materially in 2026. Amazon Bedrock AgentCore is now the primary platform for building, deploying and operating agents with managed runtime, identity, gateways, memory, browser and code execution capabilities. AWS also added generally available Web Search for AgentCore in June 2026 so agents can retrieve current web information from within AWS-managed infrastructure.

The older Amazon Bedrock Agents service has been renamed Amazon Bedrock Agents Classic and is no longer open to new customers. Existing customers can continue using it, but new projects should evaluate AgentCore rather than design around the older Agents Classic architecture. This distinction matters when reading older tutorials or comparisons.

What changed with OpenAI support on Bedrock?

OpenAI model support is now part of the current Bedrock model catalogue. AWS announced the expansion in April 2026 and now lists GPT-5.6 family models in its pricing and model documentation. The redesigned Bedrock console introduced in June 2026 is optimized for the bedrock-mantle endpoint, which supports OpenAI Responses and Chat Completions compatible APIs as well as the Anthropic Messages API.

For teams already using familiar OpenAI or Anthropic client patterns, this can reduce integration friction. It does not make Bedrock identical to using those providers directly. Model availability, release timing, supported Regions, AWS controls and Bedrock pricing should still be compared with the first-party API.

How much does Amazon Bedrock cost?

Pricing checked August 22, 2026 against AWS's official Bedrock pricing page. There is no single monthly Bedrock price. Model inference is generally usage based and varies by provider, model, Region, input tokens, output tokens, modality and service tier. Additional capabilities such as Knowledge Bases, Guardrails, model evaluation, data automation, custom model import and AgentCore can create separate charges.

Bedrock currently offers Standard, Priority, Flex and Reserved service tiers for supported models. Standard uses the normal on-demand rate. Priority charges a premium for preferential processing. Flex discounts eligible workloads that can accept less immediate processing, while Reserved is aimed at predictable capacity requirements. Batch inference is also priced below normal on-demand inference for supported models.

As one illustration of how much rates can vary, the current Sydney pricing page lists OpenAI gpt-oss-20b at $0.0721 per million input tokens and $0.3090 per million output tokens on Standard, while Flex and Batch are half those rates. This is an example, not a universal Bedrock price. Buyers should price the exact model, Region, throughput mode and supporting services their workload uses.

How does Bedrock handle enterprise data and security?

Bedrock integrates with AWS controls such as IAM, encryption, CloudTrail logging and private networking options. AWS states that Bedrock does not use customer inputs and outputs to train the underlying base models. Knowledge Bases can ground model responses in approved enterprise data, while Guardrails provide policy and safety controls that can be applied across supported models.

Security still depends on architecture. Teams need to configure permissions, data sources, network boundaries, logging, retention and application-level controls correctly. A managed AI service reduces infrastructure work but does not remove the need for governance around sensitive prompts, connected data, agent actions and third-party model providers.

What are the main limitations and trade-offs?

Bedrock can become complex because its flexibility spans many models, Regions, service tiers and auxiliary services. A team may need to understand IAM, VPC networking, CloudWatch, model quotas, cross-region inference and multiple pricing dimensions before it can operate a production workload confidently.

Model releases may also arrive on a different schedule from first-party providers, and not every model or feature is available in every Region. Applications that use provider-specific capabilities can still become coupled to one model even when the infrastructure layer is multi-model. Cost forecasting is another challenge because token usage, agent tool calls, retrieval, data processing and supporting AWS services can all contribute to the bill.

How does Amazon Bedrock compare with alternatives?

Microsoft Azure AI services can be a stronger fit for organizations centered on Azure identity, networking, data services and Microsoft enterprise agreements. Google Vertex AI can fit teams already using Google Cloud or that want Google's model and machine-learning ecosystem. Direct APIs from OpenAI, Anthropic or other model providers can be simpler when a team needs one provider and prefers the shortest route to its newest capabilities.

Bedrock is more compelling when an organization already operates on AWS, needs multiple model choices under AWS governance, or wants to combine generative AI with AWS data, security and infrastructure services. The comparison should include engineering effort, regional availability, governance, model access, latency and total workload cost rather than token price alone.

Who should choose something else instead of Bedrock?

A small team building a straightforward application around one model may not need the operational breadth of Bedrock. Going directly to the model provider can reduce the number of layers involved. Teams without AWS expertise may also prefer a simpler managed application platform if cloud architecture is not part of the problem they are trying to solve.

Bedrock makes the most sense when the AWS control plane itself is valuable: identity, networking, auditability, model choice, enterprise data integration and the ability to run AI beside existing AWS workloads. If those benefits are not important, the added platform complexity may not justify the choice.