Amazon Bedrock vs Google Vertex AI (Now Gemini Enterprise Agent Platform): Models, Pricing, RAG, Agents, Security, AI Studio, and Which to Choose in 2026

Amazon Bedrock vs Google Vertex AI comes down to three questions: which models you need, which cloud you already use, and how you want to build. Google's Gemini models are only on Google; Amazon Nova and the widest catalog of third-party models (136 models from 19 providers as of October 2026, including OpenAI's GPT-6 family) are on Bedrock; Anthropic's models, Llama and Mistral are on both. First, though, a name change most comparisons miss: on April 22, 2026, Google renamed Vertex AI to the Gemini Enterprise Agent Platform, and Model Garden, Agent Engine (now Agent Runtime) and Vertex AI Studio (now Agent Studio) were renamed with it. Prices are close for comparable models, both offer about 50% off for batch jobs, and both keep your prompts out of model training. Choose Bedrock if you run on AWS, want model choice behind one API, or need AgentCore for agents; choose Google's platform if Gemini is your model, you need long video and document understanding, or your data already lives in BigQuery. Below: the rename table, models and prices side by side, RAG, agents, security, free credits, AI Studio vs Vertex, Microsoft Foundry, and which to pick for common situations.

Jake's computer shop now runs a small warranty assistant on Amazon Bedrock. Then the web designer who looks after the shop's site mentioned that "everyone is using Gemini on Vertex now," and Jake asked the question every small business owner eventually asks: are we on the wrong one? Ethan spent an evening comparing the two properly, with the shop's actual use and actual bill in mind. This page is his comparison, written up for anyone facing the same choice.

New to these services? Our plain-English Bedrock explainer is the five-minute version of the AWS side, and the cards below cover the rest.

The short answer: a side-by-side summary

Neither is "better" in general. Each is the best place to use its own company's models, both are good places to use Anthropic's, and both are serious enterprise platforms. The most common deciding factor in practice is not the AI at all: it is which cloud your data, your team's skills and your bill already live in.

Jake: "So the designer wasn't wrong, and neither were we?"

Ethan: "Right. Gemini is excellent, and it lives on Google. Our files, our website and our bill live on AWS. That matters more than which model wins a benchmark this month."

Vertex AI is now the Gemini Enterprise Agent Platform

At Google Cloud Next on April 22, 2026, Google renamed Vertex AI to the Gemini Enterprise Agent Platform, folding Agentspace into its Gemini Enterprise product at the same time. The services did not disappear; they were renamed and grouped around building agents. If you are reading a Vertex AI tutorial from 2025, this table translates it:

Two practical notes. Existing projects keep working: API endpoints and much of the documentation still say "Vertex," which is why most people, and this page's title, still use the old name. And for new Python code, Google's current SDK is the Google Gen AI SDK (google-genai), which talks to both the developer Gemini API and the enterprise platform; the older Vertex AI generative AI SDK is the one being retired.

What each platform is

Amazon Bedrock is AWS's managed service for using foundation models through an API. You do not rent servers or GPUs; you send requests and pay per token. Around the models, Bedrock adds knowledge bases (your documents, searchable by the model), Guardrails (filters for topics, personal data and harmful content), batch jobs, model customization, and AgentCore for running agents in production. It is one of AWS's two AI services; our Bedrock vs SageMaker AI guide explains the other.

The Gemini Enterprise Agent Platform (Vertex AI) is Google Cloud's equivalent, and broader in one way: it combines what AWS splits between Bedrock and SageMaker. It serves Gemini and Model Garden's models through an API, and it also trains, tunes and deploys your own machine learning models, with pipelines, notebooks and feature stores. Since the rename, its center of gravity is agents: Agent Studio for prototyping, the open-source Agent Development Kit for code, and Agent Runtime for hosting.

So a fair one-line summary: Bedrock is AWS's API to many companies' models; Google's platform is Google Cloud's whole AI workbench, built around Gemini.

Models: who has what

Two things stand out. Gemini is the strongest reason to choose Google. Its long context and native handling of video, audio and very large documents are a real advantage for some workloads, and you cannot get Gemini on Bedrock. Breadth is Bedrock's strongest reason. Since OpenAI's GPT models arrived on Bedrock, one AWS API reaches Nova, OpenAI, Anthropic, Meta, Mistral and many open models, and switching between them is a change of model ID. Our Bedrock model list shows every model with its Regions.

Model Garden's "200+" and Bedrock's "136" are not directly comparable: Google counts open models you deploy onto your own rented machines, while Bedrock's count is models ready to call per token. For "call it by API, pay per token, no servers," the two catalogs are much closer in size than the numbers suggest.

Pricing: Bedrock vs Vertex AI, side by side

Both platforms charge per million tokens, with separate prices for input (what you send) and output (what the model writes). Comparable models land in similar ranges:

Prices are list prices in US Regions as of October 19, 2026. Note the two dated changes on Google's side: the Gemini 3.8 Flash introductory price doubles on January 1, 2027, and Gemini 4 Argon's price doubles on December 1, 2026. Any comparison made today should be redone with the new numbers before you commit.

The same model costs about the same on both. Anthropic's models are priced alike on the two clouds, so for those the choice comes down to everything else. On Bedrock, Anthropic's models cost 10% more when pinned to a geography (us., eu.) or a single Region than through the global. profile; Google also offers global and regional endpoints.

A worked example. Say a support assistant handles 100,000 requests a month, each sending 2,000 tokens and receiving 500. That is 200 million input tokens and 50 million output tokens:

The lesson is not "Haiku wins." The model tier you need moves the bill ten- or twenty-fold, while the platform moves it a few percent. Test the cheapest model that answers your questions well enough, on whichever platform you use.

The discounts. Both platforms cut about 50% for batch work that can wait hours for results. Bedrock also has a Flex tier at 50% off for work that tolerates delay and a Priority tier at 75% more for guaranteed speed, plus Provisioned Throughput for reserved capacity. Google has priority and provisioned options too. Both offer prompt or context caching, which makes repeated long prompts much cheaper; Google adds an hourly storage charge for explicit caches. Our Bedrock pricing guide works through each lever with real numbers.

Embedding models compared

Embedding models turn text (and sometimes images, audio or video) into vectors for search and RAG. They are cheap per call but matter a lot, because switching later means re-indexing everything.

For plain English text search, the cheapest option on either cloud is good enough for most document sets, and the quality of your chunking and your prompts matters more than the embedding model. For multilingual content or mixed media, test two candidates on your own documents. Whatever you pick, use the same model for indexing and for search, forever or until you rebuild the index.

Voice and real-time conversations

Both clouds now offer models that listen and speak in real time, for phone lines, voice assistants and live translation.

• Bedrock has Nova 2 Sonic, a speech-to-speech model that hears a caller and answers in a natural voice with low delay, and can call tools mid-conversation, for example to look up an order.
• Google has the Gemini Live family, which takes audio and video in and speaks back, plus dedicated live translation and transcription models, all priced per token or per minute of audio.

Voice is an area where the "use your existing cloud" rule weakens slightly: quality differences between voices and languages are easy to hear, so it is worth trying both with your own callers' accents and vocabulary before you commit.

Model versions, release notes and retirements

Models on both platforms are retired on a schedule, usually a year or more after release, and the replacement is rarely a drop-in copy. Planning for it is part of choosing a platform.

• Bedrock marks each model as Active, then Legacy, then end of life, with dates you can read through the API (legacyTime, publicExtendedAccessTime and endOfLifeTime on the model's lifecycle details). Our Bedrock model list tracks these dates, including recent retirements such as Nova Canvas.
• Google publishes model versions with release and retirement dates, and its release notes, still published under the Vertex AI name in many places, list new models and deprecations week by week. Pricing changes are announced ahead too, like the Gemini 3.8 Flash increase on January 1, 2027.

On either side, the practical defense is the same: keep model IDs in configuration rather than scattered through code, keep your twenty-question test set, and re-run it whenever you switch versions. Then a retirement notice is a one-hour task, not a crisis.

Documents and RAG: Knowledge Bases vs RAG Engine and Agent Search

Answering questions from your own documents, called retrieval-augmented generation (RAG), is the most common business use of either platform.

• Bedrock Knowledge Bases read files from S3, SharePoint, Confluence, Salesforce and the web, split and embed them, store the vectors (in a managed store or your own OpenSearch, Aurora or other database), and answer with citations through the Retrieve and RetrieveAndGenerate APIs. Our knowledge base guide covers setup and cost.
• Google's platform offers RAG Engine for building retrieval into Gemini apps, Agent Search (formerly Vertex AI Search) for search over websites and documents, and Agent Retrieval (formerly Vector Search) for your own vectors. Gemini can also ground answers in Google Search results directly, which no AWS service offers.

Both work well for a typical company document set. Google's advantages are grounding with Google Search and Gemini's very long context, which can sometimes skip retrieval entirely by reading whole document sets at once. Bedrock's are its connectors to common business sources and the choice of vector store. The bigger factor is where the documents already are: Knowledge Bases read S3 natively, Google's tools read Cloud Storage and BigQuery natively, and moving gigabytes between clouds costs time and data-transfer fees.

Agents: AgentCore vs Agent Runtime and ADK

Both companies now treat agents, models that use tools and take actions, as the main event.

• AWS offers AgentCore, a platform for running agent code you write in any framework (LangGraph, CrewAI, Strands, LlamaIndex, Google's own ADK, the OpenAI Agents SDK), with models inside or outside Bedrock. Each session runs in its own isolated microVM, and managed memory, identity, a tool gateway and observability come with it. The older Bedrock Agents remain for agents configured in the console without code. Our AgentCore explainer covers each piece.
• Google offers the open-source Agent Development Kit (ADK) for writing agents in code, Agent Studio for building and testing them visually, and Agent Runtime (formerly Agent Engine) for hosting them, with sessions, memory and evaluation. Google also champions the A2A protocol, which lets agents from different vendors talk to each other.

"Vertex AI vs ADK" is a common search, and they are not alternatives: ADK is the toolkit you write an agent with; the platform is where you can run it. ADK agents can also run elsewhere, including on AWS's AgentCore, and AgentCore can call Gemini. So the agent decision is less "AWS or Google" than "which framework, and where do I want to operate it?" Teams that write agents with LangChain can move between both clouds most easily; our LangChain with Bedrock guide shows the AWS side.

Jake: "So we could even use Google's agent kit on Amazon?"

Ethan: "We could. The frameworks don't care whose cloud they're on. What decides it is where the data and the bill are."

Security, privacy and compliance

Both are built for regulated companies, and on the fundamentals they match:

• Your data is not used to train models. Neither Bedrock nor Google's platform uses your prompts and responses to train the foundation models, and on Bedrock your data is not shared with the model providers.
• Access control comes from each cloud's own identity system: AWS IAM policies on Bedrock actions and models; Google Cloud IAM roles on the platform's resources.
• Private networking: Bedrock is reachable through VPC interface endpoints (AWS PrivateLink), so traffic never touches the public internet. Google uses Private Service Connect and VPC Service Controls, which draw a security perimeter around projects to block data leaving it.
• Data residency: both let you choose Regions, and both offer global routing for capacity. On Bedrock, the eu. or us. prefixes keep processing inside a geography; Google uses regional endpoints.
• Safety filters: Bedrock Guardrails filter topics, harmful content and personal information, and check that answers stay grounded in sources. Google offers configurable safety settings on Gemini and Model Armor for screening prompts and responses.
• Compliance programs such as SOC 2, ISO 27001 and HIPAA eligibility exist on both; check the specific services you plan to use against each cloud's current list, since new features join these programs over time.

A practical tip: if your company already has a security team that knows one cloud's controls well, that knowledge is worth more than any feature difference in this list.

Developer experience: SDKs, APIs and frameworks

On AWS, you call Bedrock with any AWS SDK (boto3 for Python), usually through the Converse API, which uses one request format for every model. Bedrock also offers an OpenAI-compatible endpoint (Mantle), so code written for OpenAI's API can point at Bedrock with a new base URL and key; our Mantle guide explains when to use which. The console's playgrounds let you try models side by side before writing code.

On Google, the Google Gen AI SDK calls Gemini on both the developer API and the enterprise platform with the same code, switching by configuration. Gemini also offers OpenAI-compatible calls. Agent Studio is the place to try prompts and models in the browser.

Frameworks such as LangChain, LlamaIndex and LiteLLM support both, which is the easiest way to keep the option of switching: write against the framework, and the provider becomes a setting. Many companies run their main workload on one cloud and test the other's models through the same framework code before deciding whether a switch is worth it.

Limits and capacity: quotas, throttling and global routing

Price matters little if your requests are refused at busy times, and the two clouds handle capacity differently.

Bedrock sets quotas per model and per Region: requests per minute and tokens per minute, visible and adjustable in the Service Quotas console. When you exceed them, calls fail with a throttling error ("Too many requests, please wait before trying again"), which our Bedrock throttling guide walks through. Two tools raise your ceiling: cross-Region inference profiles (us., eu., global.) spread requests across several Regions' capacity, and Provisioned Throughput or the Reserved tier buys guaranteed capacity for steady, heavy workloads. New accounts start with lower quotas, so request increases a week or two before a launch.

Google serves current Gemini models from a shared pool of capacity rather than fixed per-project quotas for most usage, so traffic can grow without a quota request, but during peak demand some requests can be rejected with a resource-exhausted error and need a retry. For guaranteed capacity, Google sells Provisioned Throughput too, and a priority option sits between the two.

In practice, both need the same habit in your code: retry with exponential backoff on throttling errors, and cap how many requests run at once. Both clouds' SDKs and frameworks such as LangChain have retry settings built in; turn them on rather than writing your own.

Customizing models: fine-tuning on each platform

Most companies never need to fine-tune: good prompts and RAG cover the large majority of business uses. When you do, for a fixed output style or a narrow task done millions of times, the platforms differ.

• Bedrock offers fine-tuning and distillation for a set of models, including Amazon Nova and some Llama models, using your training data in S3. A customized model on Bedrock then needs Provisioned Throughput to run, a fixed hourly charge whether you use it or not, which our pricing guide warns about. Bedrock can also import custom open-weight models you trained elsewhere.
• Google's platform offers supervised tuning for selected Gemini models, plus the full machine learning toolset (managed training, pipelines, notebooks) for training your own models from scratch. Tuned Gemini models can be served per token in many cases, without a reserved-capacity commitment.

For a team that expects serious model training, Google's single platform is more convenient; on AWS that work belongs in SageMaker AI next to Bedrock. For everyone else, this row of the comparison rarely decides anything.

Try both in an afternoon: a fair test

The best comparison is your own questions, not a benchmark table. Here is the test Ethan ran, which takes an afternoon and costs a few dollars:

1. Collect 20 real questions your assistant must handle, including a few hard ones and a few it should refuse. For the shop: warranty edge cases, price lookups, a question about a part they do not stock.
2. Write down the ideal answer for each before you test, so you judge against a standard rather than against whichever model you saw first.
3. Run them on Bedrock in the console playground or with a few lines of code, with the same system prompt and the same documents pasted in as context, on two or three candidate models.
4. Run them on Google's platform in Agent Studio (or AI Studio for a quick start) with the identical prompt and context, on Gemini 3.8 Flash and Gemini 4 Argon.
5. Score the answers blind: have someone else paste them into a sheet without the model names, and mark each right, partly right or wrong.
6. Record tokens and speed for each answer; both consoles show token counts, and speed matters for anything customers wait on.
7. Multiply by your real volume with the price table above, using the post-January 2027 Gemini prices for any decision that lasts past New Year.

For the shop's questions, the mid-priced models on both clouds landed close together, and the differences that showed up were about specific warranty clauses rather than one platform being better overall. That made cost per correct answer, and the hassle of moving, the deciding factors. Your questions will produce their own pattern, which is exactly why the test is worth an afternoon.

Moving from one to the other: what actually changes

If your test does point to the other cloud, or a customer requires it, these are the pieces that change, in the order they usually bite:

1. The model calls. Bedrock's Converse API and Google's Gen AI SDK have different request shapes. If your code uses LangChain, LiteLLM or a similar framework, this is a configuration change; if it calls the SDK directly, budget a few days.
2. Prompts. A prompt tuned for one model rarely performs identically on another. Re-run your test set and adjust; this is usually the longest step.
3. Embeddings and vector indexes. Vectors from different embedding models cannot be mixed or compared. Moving from Titan Embeddings to Gemini Embedding (or the reverse) means re-embedding every document and rebuilding the index.
4. Safety settings. Bedrock Guardrails policies and Google's safety settings and Model Armor rules are configured differently; recreate them, then test the questions that should be refused.
5. Permissions and networking. IAM policies, private endpoints and service accounts are rebuilt in the other cloud's terms.
6. Quotas. Request capacity on the new side before you move traffic.
7. Monitoring and cost alerts. Recreate budgets and dashboards so the first bill is not a surprise.

This list is also the argument for a framework from day one: steps 1 and, with care, 3 are much smaller when your application never talks to a cloud SDK directly.

For a small business: the practical view

For a shop like Jake's, the comparison shrinks to a few questions. Where are the files? Who sets things up and fixes them when they break? Which bill is already being paid? An AI assistant costs a few dollars a month in tokens either way; the hours spent learning a second cloud's consoles, permissions and billing cost more than any price difference between comparable models.

So the practical rule is simple: start on the cloud you already use, pick the cheapest model that passes your own twenty-question test, set a budget alert, and revisit the choice once a year or when a specific need appears, such as video analysis, that the other platform clearly does better. If you use neither cloud yet, pick the one whose console you find easier to understand in the first hour; you will spend many more hours in it.

Vertex AI vs Google AI Studio (and Gemini for free)

Google has two front doors to Gemini, and searchers often mix them up:

On the Gemini API's free tier, Google may use your prompts to improve its products; on the paid tier and on the enterprise platform, it does not. So AI Studio's free tier is excellent for learning and prototypes, and a poor home for customer data. AWS has no separate developer front door: Bedrock is the one service, used the same way by a hobbyist and a bank.

Free credits and trying each platform cheaply

• AWS: new accounts get up to $200 in credits under the current Free Tier ($100 at sign-up and up to $100 more for completing getting-started activities), usable on Bedrock among other services. There is no permanent free tier for Bedrock models, but small experiments cost cents.
• Google Cloud: new customers get $300 in credits for 90 days, usable on the Agent Platform. For Gemini specifically, the AI Studio free tier above lets you experiment without a Cloud project at all.

Either way, set a budget alert before you start: AWS Budgets on one side, Cloud Billing budgets on the other. An agent stuck in a loop can spend more in an afternoon than a month of normal use.

What about Microsoft Foundry (Azure)?

The third big option, which comparisons increasingly include, is Microsoft's: Microsoft Foundry, the platform formerly called Azure AI Foundry. It is the long-standing enterprise home of OpenAI's models on Azure, carries many other providers' models, and fits naturally for companies built on Microsoft 365, Entra ID and Azure. If your organization runs on Microsoft, the same logic as above applies: the cloud you already trust and pay is usually the right starting point, and the model you need is the only thing that should override it.

A rough rule for the three-way question: Gemini points to Google, Nova and the broadest single API point to AWS, and deep Microsoft 365 integration points to Microsoft. Anthropic's and the open models are available on all three, so they rarely decide it.

Which should you choose? A decision guide

And if the honest answer is "both," that is common too: one primary cloud for production, the other's models tested through a framework such as LangChain, and a switch only when a model is clearly better for a specific job. The cost of running both is mostly operational: two consoles, two bills, two sets of permissions.

Jake: "So, are we on the wrong one?"

Ethan: "No. If a customer ever needs us to analyze long repair videos, I'll try Gemini for that one job. Everything else stays where our files are."

Frequently asked questions

Ethan's verdict fit on one line of the shop's whiteboard: "Stay on Bedrock; try Gemini for video if we ever need it." The comparison had not found a winner, because there is not one in general. It had found the right answer for a shop whose files, website and bill already live on AWS, which is the only kind of answer that matters. Jake erased "Are we on the wrong one?" and wrote "No" underneath, for good measure.

Revision note. Written October 19, 2026, with prices as of that date, including Google's scheduled increases in December 2026 and January 2027. The best platform is usually the one your data does not have to travel to.