Amazon Bedrock
by Amazon Web Services
The platform for building generative AI applications and agents at production scale
Amazon Bedrock is AWS's fully managed service for building generative AI applications and agents on foundation models from thirteen-plus providers through one API, with IAM-based access control, managed RAG, guardrails and agent runtime. It is aimed at enterprises that want frontier models inside their existing AWS security, networking and billing perimeter.
Amazon Bedrock is AWS's fully managed service for building generative AI applications and agents on foundation models accessed through a single API, and AWS reports it now powers more than 100,000 organisations. Models come from AI21 Labs, Amazon, Anthropic, Cohere, DeepSeek, Luma AI, Meta, Mistral AI, OpenAI, Stability AI, TwelveLabs and Writer, so a team can switch providers without changing its integration layer. Around the models sit the parts enterprises actually buy it for: Knowledge Bases for managed retrieval-augmented generation, Guardrails for content and sensitive-information filtering plus automated reasoning checks, Flows for orchestration, model customisation and distillation, evaluations, batch inference, Provisioned Throughput for committed capacity, and a model marketplace. AgentCore, generally available since October 2025, is the model-agnostic agent runtime providing managed compute, identity and access controls for agents built in any framework. In April 2026 AWS added OpenAI models, Codex and Bedrock Managed Agents powered by OpenAI — an opinionated managed runtime that sits on AgentCore as its default compute environment. The commercial case is that Bedrock inherits the AWS control plane: access is governed through IAM roles rather than rotated API keys, traffic can stay private over PrivateLink, content is encrypted and stored at rest in the region of use, and AWS states customer inputs and outputs are neither used to improve base models nor shared with model providers. Bedrock is in scope for SOC 1/2/3, ISO 9001, 27001, 27017, 27018, 27701, 22301 and 20000, CSA STAR Level 2 and FedRAMP Moderate, and is HIPAA eligible.
A platform or ML engineering lead at an enterprise already standardised on AWS, who needs frontier models under existing IAM policy, VPC networking, regional data residency and consolidated billing rather than managing separate vendor API keys and contracts.
Frontier models from a dozen providers reachable through one API inside the AWS security and compliance perimeter you have already audited.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based, Subscription
- Target Market
- CIOs, CTOs, Enterprise Developers, ML Engineers, Platform Engineering Leads, Data Scientists
- Deployment
- Cloud-only, API-based, Multi-cloud
- Founded
- 2023
- Headquarters
- Seattle, United States
- Team Size
- 500+
- Customers
- More than 100,000 organisations, per AWS (2026); named references include Robinhood, which scaled from 500M to 5B tokens daily in six months while cutting AI costs 80%, and Epsilon
Key Features
- ✓Multi-provider model access through one API
Models from AI21, Amazon, Anthropic, Cohere, DeepSeek, Luma, Meta, Mistral, OpenAI, Stability, TwelveLabs and Writer behind a single integration, so switching providers is a config change.
- ✓AgentCore agent runtime
Generally available since October 2025, it runs agents built in any framework with managed compute, automatic authentication and access controls and no infrastructure to operate.
- ✓Bedrock Managed Agents powered by OpenAI
An opinionated managed runtime added in April 2026 pairing OpenAI frontier models and agent harness with AWS infrastructure, using AgentCore as its default compute environment.
- ✓Knowledge Bases for managed RAG
Handles parsing, chunking, embedding and retrieval as a managed service, billed per gigabyte of index storage and per thousand retrieval calls.
- ✓Guardrails with automated reasoning
Content filters, sensitive-information filters and automated reasoning checks billed per thousand text units; AWS claims blocking of up to 88% of harmful content.
- ✓Model customisation and distillation
Fine-tune on your own data and distil to smaller models, which AWS says run up to 500% faster at up to 75% lower cost than the teacher model.
- ✓IAM, PrivateLink and regional data residency
Access controlled by IAM roles rather than rotated API keys, private connectivity from your VPC, and content encrypted and stored in the region of use.
- ✓Intelligent Prompt Routing and Batch inference
Routes requests to cheaper models where quality allows for up to 30% savings, and runs non-urgent workloads at half the on-demand token price.
Capabilities
Use Cases
- •Enterprise RAG over internal documents
Stand up retrieval-augmented generation against private corpora using Knowledge Bases without building an embedding and retrieval pipeline first.
- •Production agent deployment
Run agents on AgentCore with managed identity and access control instead of operating agent compute and auth infrastructure yourself.
- •Regulated-industry generative AI
Deploy model workloads under FedRAMP Moderate, HIPAA eligibility and in-region data residency with private VPC connectivity throughout.
- •Multi-model cost and quality optimisation
Route requests across providers by price and capability, using prompt routing and distillation to cut inference spend without changing application code.
- •High-volume batch inference
Process large non-time-sensitive workloads such as document classification or enrichment at half the on-demand token price.
- •Model evaluation before production rollout
Compare candidate models on your own data with built-in evaluations rather than standing up a separate benchmarking harness.
Ideal For
Best For
- ✓AWS-standardised enterprises that need model access governed by IAM roles and audited like any other AWS service
- ✓Regulated workloads requiring FedRAMP Moderate, HIPAA eligibility, ISO 27001 and in-region data residency for model traffic
- ✓Multi-model architectures where the ability to swap between Anthropic, OpenAI, Meta, Mistral and Amazon models without re-integrating is the point
- ✓Agent deployments needing a managed runtime with built-in identity and access control rather than self-managed agent infrastructure
- ✓Teams that want private model traffic over PrivateLink with no exposure to the public internet
- ✓Cost-sensitive batch workloads, which run at half the on-demand token price
Not Ideal For
- ✗Teams that need day-one access to a model provider's newest release, since availability on Bedrock lags the provider's own API and varies by region
- ✗Projects that need to go from prototype to production fast: default service quotas are low, raising them requires a manual support ticket and justification, and requests can be denied if you are not already consuming your existing allocation
- ✗Latency-sensitive multi-step agent workflows, where practitioners report noisy-neighbour variance compounding across every hop
- ✗Advanced RAG requirements — Knowledge Bases locks you into fixed chunking strategies and vector stores, and teams needing more control end up rebuilding on OpenSearch or Pinecone anyway
- ✗Buyers wanting gateway features like semantic caching, automatic model fallback or granular per-user cost controls, which Bedrock does not provide
- ✗Organisations not on AWS, for whom the main differentiator — inheriting the AWS control plane — is worth nothing
Integrations
Deployment
Market & Ratings
More than 100,000 organisations, per AWS (2026); named references include Robinhood, which scaled from 500M to 5B tokens daily in six months while cutting AI costs 80%, and Epsilon
Market Analysis
Pros
- ✓Model choice is genuinely broad — thirteen-plus providers including OpenAI, Anthropic, Meta, Mistral and DeepSeek behind one API — so provider risk is a config change rather than a rewrite
- ✓IAM-based access control removes API key rotation entirely and produces an audit trail enterprises already know how to review
- ✓Compliance coverage is unusually complete for an AI platform: SOC 1/2/3, seven ISO standards, CSA STAR Level 2, FedRAMP Moderate and HIPAA eligibility
- ✓AWS states plainly that customer content is not used to improve base models and is not shared with model providers, and that data stays encrypted in the region of use
- ✓Cross-region inference profiles provide automatic failover when a region hits capacity constraints
- ✓Real first-party cost levers — batch at 50%, distillation, intelligent prompt routing — rather than requiring a third-party gateway to control spend
Cons
- ✗Default service quotas are low enough to block a production launch, and raising them is a manual support ticket with back-and-forth justification; AWS gives priority to customers already consuming their existing allocation, so new workloads can be denied
- ✗Latency is variable — practitioners report noisy-neighbour effects where the same prompt takes seconds longer at peak, which compounds across every step of an agent workflow
- ✗Knowledge Bases is effectively a black box with fixed chunking strategies and vector stores, and teams needing advanced retrieval end up rebuilding on OpenSearch or Pinecone
- ✗No semantic caching, automatic model fallback or granular per-user cost controls — Bedrock solves model access, not model operations
- ✗Observability is thin: CloudWatch gives basic metrics, and LLM-specific insight like cost-per-conversation or token usage by user needs external tooling
- ✗Model and feature availability lags the provider's own API and differs by region, so 'available on Bedrock' and 'available' are not the same date
- ✗Analysts flag a lock-in tension: if OpenAI-powered Managed Agents grow faster than open AgentCore adoption, the practical centre of gravity shifts to model-specific runtimes and the open-layer claim weakens
Pricing
On-Demand
Contact for pricing
- ✓Pay per input and output token
- ✓Rates vary by model provider, model size and AWS region
- ✓No commitment
- ✓Published per-model rate card on the AWS pricing page
Batch inference
Contact for pricing
- ✓50% lower price than on-demand inference for supported models
- ✓For non-time-sensitive workloads
- ✓Same model catalogue
Provisioned Throughput
Contact for pricing
- ✓Commit to model units for a 1-month or 6-month term
- ✓Discounted hourly rate
- ✓Predictable capacity for consistent traffic
- ✓Required for deploying some custom models
Model customisation
Contact for pricing
- ✓Fine-tuning charged per token or per training hour
- ✓Custom model storage at $1.95 per month
- ✓Deployment via provisioned throughput or on-demand
Bedrock is metered, not seat-licensed: you pay per input and output token at rates that vary by model, model size and region, all published on the AWS pricing page. Batch inference runs at 50% of on-demand for supported models, and Provisioned Throughput trades a 1-month or 6-month model-unit commitment for a discounted hourly rate. The platform features bill separately and add up — Guardrails charges roughly $0.15 per 1,000 text units for content filters, $0.10 for sensitive-information filters and $0.17 for automated reasoning, while Knowledge Bases charges about $5 per GB per month of index storage plus $1 per 1,000 retrieval calls and $4 per 1,000 agentic retrievals. Fine-tuning is charged per token or training hour with custom model storage at $1.95 a month. AgentCore adds the underlying model inference cost when agents invoke a model. There is no Bedrock-specific free allocation documented, and practitioners consistently report that modelling total cost across these modes is harder than the headline per-token rate suggests.
Security & Compliance
Sources
This page was written from 6 sources, 3 on domains other than aws.amazon.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Arcee Trinity
US-built open-weight model family, from on-device Trinity Nano to the 400B-parameter Trinity Large, that you can run on your own infrastructure
Abacus.AI Smaug
Open-weight Smaug Agentic, Flash and Mini models fine-tuned for long-running enterprise AI agents
Deep Cogito Cogito v2.1
MIT-licensed 671B hybrid-reasoning open model with short reasoning chains, plus custom post-training on enterprise data
GPT-6 Astra
OpenAI's frontier model for autonomous computer use, gated cyber capability and long-horizon coding