Amazon Bedrock

What it is

Amazon Bedrock is AWS’s managed service for calling foundation models from inside your own AWS account, and since 2025 it has been two products under one name. Bedrock proper is the model layer: Claude (Fable, Opus, Sonnet and Haiku), OpenAI’s GPT-5.5 and GPT-5.4 (generally available on Bedrock since 1 June 2026), Amazon Nova, gpt-oss, Llama, Mistral and a long tail of open-weight models, all behind IAM, CloudTrail and your VPC. Around it sit Guardrails (content, PII and grounding filters), Knowledge Bases (managed RAG) and batch and tiered inference. Bedrock AgentCore is the agent layer: a Runtime that runs each agent session in its own microVM, a Gateway that turns APIs and Lambda functions into MCP tools, plus Memory, Identity, Policy, Observability, Evaluations, a managed Browser and a Code Interpreter.

2026 moved the centre of gravity to AgentCore. A managed harness entered preview on 22 April 2026: you declare a model, a system prompt and tools, and AWS runs the agent loop for you, with no orchestration code, and you can export it to Strands code when you outgrow it. The same release shipped an AgentCore CLI for infrastructure-as-code deploys and AgentCore skills for coding assistants. AgentCore added US West (N. California) and Asia Pacific (Hyderabad) on 27 August. Meanwhile the original Bedrock Agents was renamed Bedrock Agents Classic and closed to new customers on 30 July 2026; AWS’s own guidance is to migrate those workloads to AgentCore.

Why ops teams end up here

The question is almost always the same: our security team says models have to run in our AWS tenancy, so is Bedrock plus AgentCore enough, or do we still need Dust, Glean or Copilot Studio? Bedrock answers the first half cleanly. Claude on Bedrock runs on AWS-managed infrastructure with zero operator access (Anthropic staff cannot reach the inference stack), prompts are not used to train models, and every call shows up in CloudTrail under your account. Usage also counts toward an existing AWS commit, which is often what gets the budget approved.

It does not answer the second half. Bedrock is a builder’s platform. Knowledge Bases can ingest S3, SharePoint, Confluence and Salesforce, but there is no chat app for employees, no search that respects each user’s permissions across company tools, and no admin console a RevOps manager can run. The ops cases where it earns its place are ones an engineer will own: a contract-intake pipeline for legal ops that classifies and extracts clause data from documents landing in S3; a pipeline-hygiene agent that reads Salesforce through a Gateway tool and writes proposed fixes to a review queue; a recruiting screen that scores applications against a rubric and must keep candidate data in the EU; the AI layer behind Amazon Connect for support.

Pricing reality

There is no seat price. You pay per token for models and per unit for everything else, billed to your AWS account.

• Claude Sonnet 5.5: $2 per million input tokens and $10 per million output tokens on the global endpoint, the same as Anthropic’s list price. Opus 5.5 is $4/$20; Haiku 5.5 is $0.10/$0.50 for prompts up to 100,000 tokens.
• Regional endpoints cost 10% more. Pinning inference to one region or a geography (US, EU, JP or AU inference profiles) for data residency adds 10% to the token bill: Sonnet 5.5 becomes $2.20/$11.
• Discounts for patience: Batch inference is 50% off on-demand for the models that offer it, the Flex tier is 50% off and Priority is a 75% premium.
• Guardrails: $0.15 per 1,000 text units for content filters or denied topics, $0.10 for sensitive-information (PII) filters.
• AgentCore: Runtime $0.0895 per vCPU-hour and $0.00945 per GB-hour, billed per second; Gateway $0.005 per 1,000 tool calls; long-term Memory $0.75 per 1,000 records a month plus $0.50 per 1,000 retrievals. The managed harness, CLI and skills carry no extra charge. Observability bills as CloudWatch.

A worked band: an internal RevOps agent handling 3,000 sessions a month, each about 20,000 input and 1,500 output tokens on Sonnet 5.5, pinned to the EU. Tokens: about $165 at global rates, about $182 with the regional premium. Runtime, generously assuming one vCPU and 2 GB held for five minutes per session: about $27. Gateway, Guardrails and Memory together: under $10. Total: roughly $215 a month, about $2,600 a year. The AWS bill is the small number. The real cost is the engineer who builds, evaluates and maintains the agent, which is why a 50-person ops team usually gets a better deal from a seat-priced product.

Best for

A platform or RevOps engineering team at a company already standardised on AWS, with at least one engineer to own the build, that needs agents or document pipelines whose data must stay inside the AWS account and region. It wins outright when the workload is high-volume and back-office (extraction, classification, enrichment, scoring), because per-token pricing on Haiku or Sonnet beats per-seat or per-conversation pricing by an order of magnitude at that scale.

Not for

An ops team that wants employees chatting with company knowledge next week. Bedrock gives you no user interface, no permission-aware search across Slack, Drive and SharePoint, and no admin console for non-engineers. Buy Dust, Glean or Langdock for that, and point them at Bedrock models if your security team insists. It is also the wrong buy for a Microsoft 365 shop whose agents live in Teams: Copilot Studio gets identity and distribution for free there.

Versus the alternatives

• Microsoft Foundry (Azure) — the market-share leader alongside Bedrock among hyperscaler model platforms. Pick it when your identity and data live in Entra and Azure, or when the agents ship into Teams through Copilot Studio. Since June 2026 the OpenAI-only advantage is gone: GPT-5.5 runs on Bedrock too.
• Google Cloud Vertex AI — the third hyperscaler option. Pick it when the warehouse is BigQuery and the company runs on Google Workspace, and when Gemini is the model you want first.
• Claude Platform on AWS — the newest entrant and the fastest-moving on features. It is Anthropic-operated and billed through AWS Marketplace, with typically same-day access to new Claude features, including Agent Skills, code execution, the Files API and Claude Managed Agents, none of which Claude on Bedrock supports. Pick it when you want Claude’s full agent toolset and can accept Anthropic as the data processor. Pick Bedrock when the requirement is literally “AWS operates it, the vendor never sees it”.
• Dust / Langdock / Glean — pick these when the buyer is an ops leader, not an engineer, and the need is assistants with connectors and permissions, not a pipeline. They complement Bedrock more than they compete with it.

Watch-outs

• Claude on Bedrock is not the full Claude API. The Bedrock version of Claude lacks the Files API, server-side web search and code execution, Agent Skills, the MCP connector, the Message Batches endpoint and Claude Managed Agents. Guard: before committing, list every Claude feature your design depends on and check it against Anthropic’s Bedrock feature table. If the list includes Skills or Managed Agents, evaluate Claude Platform on AWS instead.
• Data residency is a setting, not a default. The global endpoint is the cheapest and routes requests to any region with capacity. Guard: for EU or regulated data, call only regional endpoints or geography inference profiles, budget the 10% premium, and enforce it with an IAM policy that denies the global inference profile ARNs.
• Bedrock Agents Classic is a dead end. It is closed to new customers since 30 July 2026, and tutorials written before then still teach it. Guard: build new agents on AgentCore only, and put any existing Classic agent on a migration plan rather than extending it.
• The managed harness is still preview. It runs in four regions (N. Virginia, Oregon, Frankfurt, Sydney) and is not covered by GA commitments. Guard: use it to prototype, then export the orchestration to Strands code and deploy that on AgentCore Runtime before anything customer-facing depends on it.
• Costs spread across five meters. Tokens, Runtime, Gateway, Memory and CloudWatch all bill separately, so no single page shows what one agent costs. Guard: tag every AgentCore resource and inference profile with an agent name, activate those tags as cost-allocation tags, and set an AWS Budgets alert per agent at 120% of the forecast.

Related: the enterprise AI rollout stack and the AI agent ops stack show where a model platform like Bedrock sits next to the assistant, governance and observability layers.