What Is Amazon Bedrock? AWS's Foundation Model Platform Explained (2026)

Bedrock is not a model. It is a menu. AWS built a fully managed service that gives you API access to over 100 foundation models from 15+ providers, and you never touch a server, provision a GPU, or manage an endpoint.

That distinction matters because it changes who can build with generative AI. If your team already runs infrastructure on AWS, Bedrock turns model selection into a configuration choice rather than an infrastructure project. The same API call that hits Claude today can point to Llama tomorrow. For the 100,000+ organizations already using it, that flexibility is the point.

What Is Amazon Bedrock

Amazon Bedrock is a fully managed, serverless service that provides secure, enterprise-grade access to high-performing foundation models through a single API. It went generally available on September 28, 2023, after an initial preview announcement on April 13, 2023.

The value proposition is straightforward. You pick a model (or several models), call it through a standardized API, and AWS handles the compute, scaling, and infrastructure behind the scenes. No model weights to download. No GPU instances to configure. No cold-start debugging at 2 AM.

The model catalog spans Amazon's own Nova family, Anthropic's Claude, Meta's Llama, Mistral AI, AI21 Labs, Cohere, DeepSeek, Stability AI, and several others. The Bedrock Marketplace adds another 100+ models beyond the core catalog. This multi-provider approach is what separates Bedrock from Azure AI Foundry (which centers on OpenAI) and Google Vertex AI (which centers on Gemini). You are not locked to a single vendor's model roadmap.

Bedrock connects directly to S3, Aurora, OpenSearch, and the broader AWS ecosystem. If your data already lives in AWS, the integration path is short. For background on where Bedrock fits in the broader AI tool landscape, see the AI Tools Hub and the AWS AI Services sub-hub. Ready to build? The How to Use Amazon Bedrock guide walks through setup from zero to first API call.

How Bedrock Works

The architecture follows a pattern familiar to anyone who has used AWS managed services. You make an API call. AWS routes it to the appropriate model endpoint, handles inference compute, and returns the response. You pay per token (or per image, per second of audio, per API call) with no upfront commitment required.

APIs

Bedrock supports five API formats, which means you do not have to rewrite your application when switching models:

• Converse API: AWS's unified format that works across all Bedrock models
• Invoke API: Direct model invocation for provider-native payloads
• Messages API: Anthropic-native format for Claude models
• Chat Completions API: OpenAI-compatible format for migration from GPT-based apps
• Responses API: OpenAI-compatible format for newer OpenAI model patterns

Inference Options

Three routing modes let you balance compliance, latency, and throughput:

• In-Region: Data stays in a single AWS region. Required for strict regulatory compliance (HIPAA, data residency).
• Geo Cross-Region: Routes across regions within a geography (US, EU, or Australia) for higher throughput while keeping data within regulatory boundaries.
• Global Cross-Region: Maximum throughput with routing across all available regions.

Key Features

Bedrock is more than model hosting. AWS has layered enterprise features on top of the inference layer that handle the hard parts of production AI: retrieval, orchestration, safety, and customization.

Knowledge Bases (Managed RAG)

Fully managed retrieval-augmented generation. Connect data sources (S3, Confluence, Salesforce, SharePoint, or a web crawler), and Bedrock handles chunking, embedding, and vector storage. Supported vector stores include Aurora PostgreSQL, OpenSearch Serverless, Neptune Analytics, MongoDB Atlas, Pinecone, and Redis. GraphRAG support adds knowledge graph traversal for complex multi-hop queries. Every response includes source attribution so you can trace answers back to the original document.

Agents and AgentCore

Bedrock Agents handle multi-step task automation with tool use, code interpretation, and memory retention across sessions. Multi-agent collaboration lets you chain specialized agents together for complex workflows.

AgentCore, which reached general availability in late 2025, goes further. It is an end-to-end agentic platform with its own runtime, gateway, memory, identity management, policy engine, code interpreter, browser tool, evaluations, and observability. The key difference: AgentCore works with any framework (LangChain, CrewAI, custom) and any model (not just Bedrock models). The Managed Harness added in April 2026 handles deployment, scaling, and security for agentic workloads.

Guardrails

Six configurable safety policies: content filters with adjustable sensitivity thresholds across six categories, denied topics, word filters, PII detection and redaction, contextual grounding checks, and automated reasoning that uses formal logic for factual validation (AWS Bedrock Guardrails documentation). Two tiers (Standard and Classic) with the ApplyGuardrail API available for models running outside Bedrock. For a deeper dive into configuration options and deployment patterns, see What Is Amazon Bedrock Guardrails.

Custom Models

Four customization paths: supervised fine-tuning (SFT), reinforcement fine-tuning (RFT), continued pre-training on your own data, and model distillation. Distillation lets you compress a large model's capabilities into a smaller, cheaper one - AWS claims up to 500% faster inference and 75% lower cost with distilled models (AWS Bedrock Model Distillation documentation). You can also import custom models trained elsewhere.

Prompt Management and Optimization

Prompt caching with 5-minute and 1-hour TTL windows reduces costs by up to 90% and latency by 85% for repeated prompt prefixes (AWS Bedrock Prompt Caching documentation). Intelligent Prompt Routing automatically selects the most cost-effective model for each request - AWS claims up to 30% cost reduction from routing alone (AWS Bedrock Intelligent Prompt Routing documentation). Prompt Optimization rewrites your prompts for better model performance.

Flows

A visual workflow builder for chaining foundation models, agents, knowledge bases, and guardrails into multi-step pipelines. Drag-and-drop orchestration without writing glue code.

Data Automation

GenAI-powered processing for documents, images, audio, and video. Extracts structured data from unstructured inputs at scale.

The Nova Model Family

Amazon's first-party model lineup launched in late 2024. The Nova family covers text, image, video, and speech, giving AWS a model option at every price point without depending entirely on third-party providers.

The Nova models are priced aggressively. Nova Micro at $0.035 per million input tokens is one of the cheapest foundation models available from any provider. Nova Premier at $2.50/$12.50 (input/output) targets workloads where you need a high-capability model but do not want to pay Claude Opus or GPT-5 pricing.

Models and Pricing

Bedrock's pricing is pay-per-token with no minimum commitment on the Standard tier. The model catalog spans price points from $0.035 per million input tokens (Nova Micro) to $15.00 per million (Claude Opus 4). Verified May 2026.

Service Tiers

Additional costs: Knowledge Bases $0.001-$0.002/query. Guardrails content filters $0.15/1K text units. Batch inference 50% off on-demand. Prompt caching reduces costs up to 90%. Pricing verified May 14, 2026.

Bedrock vs the Competition

Three platforms dominate managed foundation model access. Each reflects the strategic priorities of its parent company.

Bedrock's advantage is model diversity. If you need Claude for reasoning, Llama for cost-sensitive batch processing, and Stability AI for image generation, Bedrock lets you use all three through the same API with the same security controls. Azure and Google both steer you toward their in-house flagship models.

The trade-off: if your organization is already deep in Microsoft 365 or Google Workspace, the competing platforms offer tighter integration with those specific productivity ecosystems. Bedrock's ecosystem advantage only applies if you are already on AWS.

Who Should Use Bedrock

Limitations

Bedrock solves the infrastructure problem, but it introduces its own constraints. Enterprise buyers should weigh these against the platform's strengths before committing.