Amazon Bedrock Review, Pricing & Alternatives (2026)

Amazon Bedrock is part of Amazon Web Services and became generally available in 2023. It is built for developers and enterprises that want to use large language models inside their AWS accounts, with the same identity, networking, logging and compliance controls they use for other AWS services. Data sent to Bedrock stays in the customer's AWS environment and, according to AWS, is not used to train the underlying models. Customers range from startups to banks, healthcare companies and governments, and Bedrock is a common choice for teams that already run their applications on AWS and want one bill, one security model and private network access to models.

The model catalog is Bedrock's main draw. AWS lists models from AI21 Labs, Amazon, Anthropic, Cohere, DeepSeek, Google's Gemma, Luma AI, Meta, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI, Qwen, Stability AI, TwelveLabs, Writer, xAI and Z.AI. That includes Anthropic's Claude family, Amazon's own Nova models for text, speech, images and video, Meta's Llama, OpenAI's GPT and open-weight gpt-oss models, and image models from Stability AI. Developers call them through the Converse API with the same code, so swapping models takes a parameter change. The Playground in the console lets teams compare prompts across models before writing code.

Around the models, Bedrock adds building blocks. Knowledge Bases connect models to company documents in S3 and other sources for retrieval augmented generation. Guardrails filter harmful content, block topics, redact personal data and check answers against source material. Model customization supports fine-tuning and distillation, and custom models run on provisioned throughput. Amazon Bedrock AgentCore is a set of services for running agents at scale with any framework, including Runtime microVMs, Gateway for turning APIs into MCP tools, Memory, Identity, a managed Browser tool, Code Interpreter, Web Search, observability and evaluations. Bedrock Flows and Prompt Management help teams build and version workflows.

Bedrock charges for what you use. On-demand inference bills per input and output token for text models, per image for image models and per input token for embeddings, with rates that vary by model and region. AWS says batch inference for select models costs 50% less than on-demand. Provisioned Throughput sells model units by the hour, with no commitment or one-month and six-month terms that lower the hourly rate. AgentCore has its own rates, such as $0.0895 per vCPU-hour for Runtime, Browser and Code Interpreter, $7 per 1,000 web search queries and $0.005 per 1,000 Gateway invocations. There is no flat subscription.

Bedrock's strengths are a broad model catalog under one API, AWS security, private networking and compliance, data that stays in your account, and mature tooling for RAG, guardrails and agents through AgentCore. Its weaknesses are AWS console complexity, pricing that spreads across many services, regional differences in model availability and slower access to some newest models than each lab's own API. Google Vertex AI is the Google Cloud equivalent. Microsoft Foundry serves Azure customers with OpenAI, Claude and thousands of other models. Hugging Face hosts open models. OpenRouter routes across providers with one key. Choose Bedrock if your stack already runs on AWS.