What Is Amazon Bedrock? AWS AI in Plain English (2026)

Amazon Bedrock is AWS’s service for using ready-made AI models — you send text in, pay per word-piece processed, and get the model’s answer back, with no server, no GPU, and nothing to install. And here is the part that surprises everyone on day one: Amazon’s flagship AI service sells almost no Amazon AI. Bedrock is a marketplace of other companies’ models — Anthropic’s Claude, Meta’s Llama, Mistral, DeepSeek, and as of 2026 even OpenAI models — more than sixty of them behind one AWS door, next to Amazon’s own Nova family. You are not buying Amazon’s chatbot; you are renting frontier AI by the sentence, inside your own AWS account, where your data stays put. This post explains what Bedrock actually is, what it honestly costs, how it differs from ChatGPT and from SageMaker, and the news that matters this month: Bedrock’s models can now search the live web.

Jake almost talked a small business out of AI entirely. A family-run store wanted a support assistant that could answer questions from their catalog, and Jake did the responsible thing: he priced it. A GPU machine to run a decent model came out at hundreds of dollars a month before it answered a single customer — for a store whose whole IT budget was a shared laptop. He drafted the "sorry, this is enterprise stuff" message… and deleted it after one afternoon in the Bedrock playground. No server. No GPU. A model rented per request, wired to the catalog, tested in the browser first. The first month of real usage cost less than the coffee Jake drank while building it — single-digit dollars, billed for exactly the sentences processed and not one idle hour. His notebook entry from that week: "I priced a kitchen when all they needed was to order food."

Ethan: "Bedrock is a food court with one cashier. Anthropic runs a counter, Meta runs a counter, Mistral, DeepSeek, even OpenAI now — and Amazon runs its own Nova counter in the corner. You order from any kitchen, the one cashier bills you, and you never step behind a stove. That’s the whole trick. And when someone says they need SageMaker instead — that’s the empty kitchen down the road you rent to cook your own recipes. Most people don’t want a kitchen. They want lunch."

The vending-machine reveal: what Bedrock actually is

Strip the branding and Bedrock is three promises. First: the models are already running — AWS hosts them, scales them, patches them; you call an API and get an answer, the same way S3 gives you storage without a hard drive. Second: one door, many vendors — instead of signing up with each AI company separately, juggling their keys and their bills, everything arrives through your existing AWS account, your existing IAM permissions, your existing invoice. Third, and least appreciated: your data stays yours. Prompts and responses are not used to train the models, and traffic stays inside your AWS environment — the property that lets a clinic, a bank, or a law office use frontier AI without mailing its documents to a third party’s training pile. That third promise, more than any feature, is why Bedrock exists as a business.

The Bedrock models on the shelf

The catalog moves monthly — as of August 2026 it spans more than sixty models from over a dozen companies — but the families worth knowing are stable:

Notice what that table means strategically: picking Bedrock is refusing to marry any one AI company. The model your app uses becomes a configuration line, not an architecture. When a better or cheaper model appears — and one appears roughly every quarter — you switch by changing a name, not by rebuilding around a new vendor’s API. In a field moving this fast, that switchability is worth more than any single model on the list.

Bedrock pricing: you rent sentences, not servers

Bedrock’s default billing is on-demand, per token — a token being a word-piece, roughly three-quarters of an English word. Prices are quoted per million tokens, input and output counted separately. The spread is wide and worth respecting: as of August 2026, Amazon’s smallest Nova model costs a few cents per million input tokens, mid-tier models sit around a dollar or few, and the largest frontier models run to several dollars per million — always check the live pricing page, because these numbers get revised in one direction: down. Two things follow. Experimenting is essentially free — a long chat is a few thousand tokens, so your first week of learning costs less than a snack. And production costs are a multiplication: tokens per request, times requests per day, times the model’s rate — the same compounding honesty we walked through for AI agents, where one task can quietly spend fifty model calls. Bulk jobs get a standing discount: batch mode processes big piles overnight at half the on-demand rate. One warning flag this early, expanded below: the pricing page also offers provisioned throughput — reserved capacity rented by the hour. That option is for heavy production traffic, and choosing it casually is how hobby projects generate legendary bills.

"Why not just use ChatGPT?" — the honest comparison

For chatting, you should just use a chatbot — ChatGPT, Claude, Gemini, whichever you like; Bedrock would be the expensive, awkward way to have a conversation. Bedrock is for building: the moment AI needs to live inside your product, your workflow, or your company’s rules, the calculus flips. Your app’s data stays in your AWS account instead of traveling to a consumer service. Access control rides the IAM permissions your team already manages, and spending lands on the AWS bill your finance person already reads — with Budgets and alarms standing guard like for any other AWS service. Compliance certifications come with the platform instead of being your problem. And the multi-vendor shelf means the model is swappable. None of that matters for asking a chatbot to draft an email; all of it matters the day a customer’s data flows through the prompt.

Bedrock vs SageMaker: lunch vs the rented kitchen

AWS sells two AI doors and the split is clean. Bedrock: use finished models. Somebody else trained them; you send prompts. SageMaker: build your own. Notebooks, training jobs, GPU fleets, deployment endpoints — we walked through that whole workshop in plain English. The trap is status-driven: teams reach for SageMaker because training a custom model sounds serious, then discover that a frontier model plus good instructions plus their documents beats the small custom model they could afford to train — at a fraction of the effort. The boring guidance that saves months: start at Bedrock; graduate to SageMaker only when a model must learn patterns that cannot be put in the prompt. Fine-tuning — the middle path where a foundation model gets nudged with your examples — exists on both, and even then Bedrock’s managed version is the gentler on-ramp.

What arrived this month: models that can search the web

The classic embarrassment of AI models is the knowledge cutoff — ask about anything recent and they either admit ignorance or invent confidently. Through the summer of 2026 AWS closed that gap: Bedrock now has built-in web search. It arrived first as a tool for agents on AgentCore and, as of August, sits directly in Bedrock model inference — any supported model can pull current, cited web results, with the answers grounded in sources and, notably, without your queries leaving the AWS environment for a third-party search provider. The same season made AgentCore generally available: AWS’s managed home for AI agents — the loop-running, tool-calling programs we took apart honestly in the agents post — with sessions that can run up to fourteen days and the industry converging on shared protocols for how agents talk to tools. Read the direction: Bedrock started as a shelf of models; it is becoming the place where agents live. If you learned the agent loop from our post, AgentCore is where AWS expects you to deploy it.

The toolbox around the models

Honest aside: when Bedrock is the wrong answer

Three cases, honestly. You just want AI for yourself: use the free chatbots — building through Bedrock to ask questions is renting a food court to make one sandwich. You want AI on private files without any cloud at all: that is the local-LLM road — we keep guides for running DeepSeek on a Windows PC and running Gemma locally — slower and weaker than frontier rentals, but the data literally never leaves your room. You skipped the seatbelts: Bedrock has no meaningful always-free allowance (new-account credits aside), and while on-demand tokens are honest money, two habits protect you before any real traffic:

1. Stay on on-demand billing until sustained production traffic forces the conversation. Provisioned throughput reserves model capacity by the hour — enterprise-grade, and priced like it. It should never be the checkbox a beginner discovers by accident.
2. Set a Budget alarm first — the same five-minute seatbelt from our AWS billing guide — and give agent-style workloads their own alarm, because loops multiply calls faster than intuition tracks.

Your first ten minutes, concretely

1. Sign into the AWS Console and open Amazon Bedrock (any major region works; a few models are region-picky).
2. Open Model access and enable the models you want to try — a one-time checkbox per provider, usually granted instantly.
3. Go to Playgrounds → Chat, pick a small cheap model first — a Nova or Haiku-class one — and talk to it. This browser chat is billed at the same per-token rates: pennies.
4. Ask the same three questions to a frontier-class model and compare. This teaches the real lesson of the shelf — where the price buys quality and where it doesn’t.
5. Before any real project: set the Budget alarm, then try a Knowledge Base against a folder of your own PDFs — that demo is what convinces every skeptic in the room.

FAQ — Amazon Bedrock, answered straight

What is Amazon Bedrock in one sentence?

A pay-per-use AWS service that lets your apps call dozens of ready-made AI models — Claude, GPT, Llama, Nova, and more — through one API, with no servers to run.

Whose models are on Bedrock?

Anthropic, OpenAI, Meta, Mistral, DeepSeek, Cohere, AI21, Stability, and Amazon’s own Nova line — sixty-plus models from over a dozen companies as of August 2026, and the list grows monthly.

Is ChatGPT on Bedrock?

ChatGPT the product, no — that is OpenAI’s own app. OpenAI models, yes: as of 2026 they are available through Bedrock like the other families.

How much does Bedrock cost?

Per million tokens, input and output separately: cents for small models, a few dollars for frontier ones (rates as of August 2026 — check the pricing page). A learning session costs pennies; production costs are your volume times the rate.

Is there a free tier?

No always-free allowance like S3 or Lambda have. New-account promotional credits can cover experiments — but treat every Bedrock call as billed, and set a Budget alarm on day one.

What is a token?

A word-piece — roughly three-quarters of an English word. "Unbelievable" may be three tokens; a page of text is a few hundred. Both your prompt and the model’s reply count.

Bedrock or SageMaker — which do I need?

Bedrock to use finished models; SageMaker to build or train your own. Start at Bedrock; most projects never need to graduate.

Why use Bedrock instead of a chatbot subscription?

For personal chatting, don’t. Bedrock earns its place when AI goes inside a product: data stays in your account, access rides IAM, billing joins your AWS invoice, and models stay swappable.

Does Amazon train AI on my prompts?

No — prompts and outputs are not used to train the models, and your content stays within your AWS environment. That privacy line is Bedrock’s core selling point.

Can Bedrock models search the internet now?

Yes — as of summer 2026, web search is built in: models and agents can pull current, cited web results without a third-party search provider, and without your queries leaving AWS.

What is AgentCore?

AWS’s managed platform for running AI agents — programs that loop, call tools, and act — with long-lived sessions (up to fourteen days) and managed tool connections. It is where Bedrock-built agents go to live in production.

What are Knowledge Bases?

Bedrock’s connect-your-documents feature: point it at files (S3 and other sources), and models answer questions from your content with citations — the "chatbot that knows our manuals" pattern without building search yourself.

What are Guardrails?

Configurable filters on model input and output — blocked topics, redacted personal information, content limits — applied consistently no matter which model runs underneath.

Do I need to code to try Bedrock?

Not to try it — the console Playgrounds are point-and-click chat. Building it into an app needs code, but the first afternoon doesn’t.

What is the provisioned throughput trap?

Provisioned throughput rents reserved model capacity by the hour — correct for heavy steady production, ruinous as a casual choice. Beginners stay on-demand, full stop.

Can I use Claude on AWS Bedrock?

Yes — Anthropic’s Claude family is one of Bedrock’s flagship shelves. Enable it once under Model access, and you call Claude with your AWS credentials and pay per token on your AWS bill — no separate Anthropic account needed.

Is Amazon Bedrock related to Minecraft Bedrock Edition?

Not even slightly — pure name collision. Minecraft Bedrock Edition is the game (that is where "bedrock dedicated server" searches lead); Amazon Bedrock is the AWS AI service. If you came here holding a pickaxe, no hard feelings.

Which model should a beginner start with?

A small cheap one — Nova Micro or a Haiku-class model — then compare against a frontier model on your real task. Switching is the whole point of the shelf; let the comparison, not the marketing, pick your model.

Where to go next

Revision note. Written August 25, 2026, the week AWS finished building web search into Bedrock and the agent platform went mainstream — a good week to answer the question honestly. Model lists and prices in this post are stamped "as of August 2026" on purpose folks!: this shelf changes monthly, which is, as the post argues, exactly why the shelf matters more than any model on it. The AWS series  continues — the promised first real project end-to-end is next on the workbench, and Bedrock just earned itself a station in it. If a price, a model name, or a feature here has drifted by the time you read this, tell us and we will bring the page back to true. And if you priced a GPU server and quietly shelved your AI idea: pour the coffee and open the playground — your idea was never too expensive. The kitchen was😁.