AWS Bedrock vs Google Vertex AI Compared

If your company wants to build with large language models without operating your own GPU fleet, you will almost certainly evaluate two platforms: AWS Bedrock and Google Vertex AI. Both are fully managed services that give enterprises hosted access to frontier models, retrieval pipelines, agent frameworks, and compliance certifications in one console. They are also the front line of a much bigger fight — Amazon and Google are each spending tens of billions of dollars a year on AI infrastructure, and this comparison is where that spending becomes visible to an ordinary engineering team choosing an API.

This guide walks through what each platform actually does, where the real technical and organizational differences sit, what current pricing looks like, and how to pick between them without getting distracted by marketing claims that neither company can fully back up.

What Each Platform Actually Is

Think of Bedrock and Vertex AI less as "AI products" and more as model distribution layers — the piece of cloud infrastructure that sits between a raw foundation model and your application code, handling authentication, billing, safety filtering, and data connections so you don't have to build that plumbing yourself. That framing matters because it explains why the real competition is not really "which AI is smarter" — it's "which plumbing fits the data you already have."

AWS Bedrock launched in general availability at re:Invent 2023 and has expanded every year since. As of mid-2026, Bedrock offers access to more than 60 large language models, three embedding models, and half a dozen specialized models (image, video, speech) across 42 AWS regions, sourced from 13 different model providers, according to AWS's own Bedrock documentation. Its core building blocks:

• Model marketplace — Anthropic Claude (Opus, Sonnet, Haiku), Amazon's own Nova family, Meta Llama 4, Mistral Large and Codestral, Cohere Command R and R+, AI21's Jamba, Stability AI's image models, and DeepSeek.
• Knowledge Bases — managed retrieval-augmented generation (RAG) that indexes documents sitting in an S3 bucket into a vector store automatically.
• Guardrails — configurable content filters, PII redaction, and topic blocking that sit between the model and your users.
• Bedrock Agents and AgentCore — a framework for tool-using AI agents that reached general availability in April 2026, adding policy enforcement (via AWS's Cedar authorization language), observability, and episodic memory for long-running agent workflows.
• Model evaluation — built-in benchmarking that lets you test candidate models against your own prompts and data before committing to one.

Google Vertex AI traces back to a 2021 rebrand of Google's older "AI Platform," unifying what had been a fragmented set of machine learning tools into one console. Its core building blocks:

• Models — the Gemini family (now on Gemini 3, including the Flash and Pro tiers), Imagen for image generation, Veo for video, and third-party models including Anthropic Claude (added via partnership in 2024), Meta Llama, and Mistral.
• Agent Builder / Gemini Enterprise Agent Platform — Google rebranded its "Agentspace" product to Gemini Enterprise in late 2025, then folded the developer-facing agent tooling into what its own documentation now labels the "Gemini Enterprise Agent Platform (formerly Vertex AI)," per Google Cloud's product page.
• Full MLOps pipeline — Workbench notebooks, Vertex Pipelines, and a Model Registry that support classical machine learning, not just generative AI. This is the biggest structural difference between the two platforms, and we'll come back to it.
• Vector Search — a native, high-scale vector database for RAG applications.
• BigQuery integration — you can call a Gemini model directly from a SQL query using the ML.GENERATE_TEXT function, without writing any application code at all.

The single most consequential difference between these platforms is not which chatbot is smarter — it's where your company's data already lives. If your data lake is in S3, Bedrock's Knowledge Bases turn retrieval-augmented generation into a nearly turnkey setup. If your analytics run on BigQuery, Vertex lets you call a frontier model from inside a SQL statement. That's usually the deciding factor, not the model leaderboard.

Head-to-Head Comparison

Table 1. AWS Bedrock vs Google Vertex AI, feature by feature (accurate as of July 2026).

Dimension	AWS Bedrock	Google Vertex AI
Flagship models	Claude Opus/Sonnet/Haiku (Anthropic), Amazon Nova (Micro, Lite, Pro, Premier)	Gemini 3 (Flash, Pro), Claude (via partnership)
Model provider count	13 providers, 60+ LLMs	Primarily Google first-party, plus Anthropic, Meta, Mistral
Native cloud data store	Amazon S3, RDS, OpenSearch	BigQuery, Cloud Storage
Managed RAG	Knowledge Bases (points at an S3 bucket, auto-indexes)	Vector Search + Agent Builder
Agent framework	Bedrock Agents / AgentCore (GA April 2026)	Gemini Enterprise Agent Platform (formerly Agentspace)
Classical ML / custom training	Limited — mostly fine-tuning of hosted models	Full pipeline: Workbench, Pipelines, Model Registry
Query models from SQL	Not natively	Yes, via BigQuery ML.GENERATE_TEXT
Compliance	SOC 2, HIPAA-eligible, FedRAMP High	SOC 2, ISO 27001, HIPAA-eligible (under BAA), FedRAMP High/Moderate depending on service
Billing models	On-demand, provisioned throughput, batch (50% discount)	Pay-per-token, prepay/postpay plans (introduced March 2026), committed use discounts
SDK languages	Python, JavaScript, Java, Go, .NET, Ruby	Python (deepest support), JavaScript, Java, Go
Speech and video	Nova Sonic (speech-to-speech, GA early 2026), Nova Reel (video)	Veo (video), native Gemini multimodal audio
Best fit	AWS-native shops, multi-model strategy, agent governance	GCP shops, BigQuery-heavy analytics, Gemini-first teams, classical ML alongside GenAI

Models: Where the Real Product Differences Live

For most engineering teams, model access is the first filter, and it is worth being specific rather than vague about who has what.

If you want Claude, both platforms offer it, but the relationship is not symmetric. Amazon has invested billions of dollars directly in Anthropic, and new Claude releases have consistently landed on Bedrock at or near the same time Anthropic ships them directly, per Amazon's own investment disclosures and Anthropic's model release notes. Vertex AI added Claude through a separate partnership in 2024 and generally receives new Claude models on a similar but not identical timeline.

If you want Gemini, Vertex AI is the only enterprise cloud option — Gemini is Google's first-party model family, and it does not run on Bedrock or Azure. Gemini 3 Flash is explicitly positioned by Google as a frontier-reasoning model at flash-tier latency and cost, aimed at high-volume agentic workloads rather than one-off chat queries.

If you want an open-weight model like Llama 4 or Mistral Large, both platforms host them, and pricing tends to be close because the model weights themselves are identical — the cloud is just charging for compute and hosting, not for the model's intelligence. Teams weighing other ways to run open models can also compare LLM hosting options such as Modal, Replicate, and RunPod.

If you want OpenAI's GPT family, neither Bedrock nor Vertex AI offers it. That is Microsoft Azure AI Foundry's exclusive lane, a direct consequence of Microsoft's investment in OpenAI. This is a genuinely hard constraint, not a negotiating position — if your organization has standardized prompts and evals around GPT-5-class models, moving to Bedrock or Vertex means re-testing against a different model family, not just swapping an API key.

Is Bedrock or Vertex AI cheaper for the same model?

For identical open-weight models like Llama or Mistral, token prices on Bedrock and Vertex AI typically sit within roughly 10% of each other, because both clouds are pricing the same underlying compute rather than a proprietary model. For first-party models — Claude and Nova on Bedrock, Gemini on Vertex — direct comparison is less meaningful because the models themselves are not interchangeable.

Pricing: What Actually Moved in 2026

Pricing on both platforms changes often enough that any specific number in this article should be treated as a snapshot, not a permanent fact — always check the live pricing page before committing budget. That caveat aside, a few concrete data points from mid-2026 are worth knowing because they illustrate how aggressively both companies are competing on price.

On Bedrock, Anthropic's Claude Sonnet 5 launched with promotional pricing of $2 per million input tokens and $10 per million output tokens, a discount that AWS structured to run through August 31, 2026, after which standard pricing of $3/$15 per million tokens takes over, according to AWS Bedrock's pricing page. The older Claude Sonnet 4.6 saw its price cut roughly in half, from $3/$15 down to $1.50/$7.50 per million tokens. Amazon's own Nova Pro model dropped from $0.80/$3.20 to $0.40/$1.60 per million tokens in the same window. Amazon Nova Micro, the smallest model in the family, prices at roughly $0.035 per million input tokens and $0.14 per million output tokens, making it one of the cheapest hosted models available on any major cloud.

On Vertex AI, Google restructured Gemini pricing at Google I/O 2026. Gemini 3.5 Flash launched at $1.50 per million input tokens and $9.00 per million output tokens — priced roughly 25% below the older Gemini 3.1 Pro while outperforming it on coding benchmarks, according to Google's release notes. A subsequent update, Gemini 3.6 Flash, landed on July 21, 2026 at $1.50 per million input tokens and $7.50 per million output tokens, a further 17% cut to output pricing relative to the 3.5 Flash generation it replaced. Google also introduced a structural billing change: Prepay and Postpay plans replaced the older pure pay-as-you-go model starting March 23, 2026, per Google's Gemini API pricing documentation.

Both platforms also offer non-real-time discount tiers. Bedrock's batch inference mode processes requests within a 24-hour service-level agreement at roughly a 50% discount versus on-demand pricing — useful for large offline jobs like bulk document summarization where latency doesn't matter. Vertex AI offers committed-use discounts and, as of February 2026, introduced separate billing for its Agent Engine's Sessions and Memory Bank features, with a free tier covering 50 vCPU-hours and 100 GB-hours of memory per month.

Which platform has cheaper token pricing overall?

Neither platform is categorically cheaper. Both AWS and Google have cut prices aggressively through 2026, often within weeks of each other, and headline token prices now differ more by model tier (small vs. flagship) than by cloud provider. Total cost of ownership is usually driven more by data egress fees, vector storage, and engineering time spent on integration than by the per-token rate itself.

Data Integration: The Real Source of Lock-In

This is the section that actually determines most enterprise decisions, more than any model benchmark does.

Bedrock plus S3. If your organization's document repository already lives in Amazon S3, Bedrock Knowledge Bases make retrieval-augmented generation close to turnkey. You point Bedrock at a bucket, it chunks and embeds the documents, and it builds a managed vector index automatically. There is very little custom pipeline code to write.

Vertex plus BigQuery. If your analytics workloads already live in BigQuery, you can call Gemini directly from a SQL statement using ML.GENERATE_TEXT, without writing any application code at all. For a data analytics team that already thinks in SQL, this collapses what would otherwise be a multi-step integration project into a single query.

Think of it like grocery shopping near your house versus driving across town: the AI platform features are the same rough quality either way, but one route has a much shorter distance between where your data sits and where the model needs it. Most enterprises already have the bulk of their structured or unstructured data committed to one cloud provider by the time they start evaluating AI platforms, and switching that underlying data infrastructure is a far bigger project than switching an AI vendor. That is usually the deciding factor, not the AI platform's feature list.

Agent Frameworks: Bedrock AgentCore vs Gemini Enterprise

Both companies made agent tooling their headline 2025-2026 investment, and both frameworks are aimed at a similar problem: letting an LLM take multi-step actions (call a tool, query a database, trigger a workflow) safely and observably, rather than just answering a single prompt.

Bedrock AgentCore reached general availability in April 2026, following its announcement at re:Invent 2025. It uses AWS's Cedar policy language — the same authorization system that powers AWS Verified Permissions — to define, before an agent runs, exactly which tools and actions it is allowed to invoke. AgentCore also adds continuous evaluation of agent behavior (helpfulness, correctness, safety, tool usage) and "episodic memory" designed to let agents persist context across long-running, multi-session workflows rather than starting fresh every conversation.

Google's Gemini Enterprise Agent Platform (the current name for what was Vertex AI Agent Builder, then briefly Agentspace) takes a similar shape: build conversational or search-oriented agents, wire them into internal tools and data sources, and deploy them with access controls. Because it sits inside the broader Vertex AI umbrella, it inherits the same BigQuery and Google Workspace integrations that make Vertex attractive for GCP-native teams.

The practical difference is philosophical as much as technical. AgentCore is built around explicit, pre-declared permission policies — you write down what an agent can do before it runs, and the platform enforces it deterministically. Gemini Enterprise leans more on Google's existing identity and access management layer plus its own guardrails. Neither approach is objectively safer; they reflect each company's broader security architecture.

What is Amazon Bedrock AgentCore used for?

AgentCore is Amazon's framework for building, deploying, and governing AI agents that take multi-step actions rather than answering single prompts. It enforces which tools an agent may call through Cedar-based policies, tracks agent behavior for safety and correctness, and gives agents persistent memory across sessions. It reached general availability in April 2026.

Developer Experience and Team Fit

Both platforms have mature SDKs in the major languages and reasonably complete documentation, but they feel different to work in day to day, and that difference correlates with what kind of team you have.

Bedrock feels more like a model marketplace: you pick a model, call an API endpoint, and integrate. The mental model is closer to "swap in whichever LLM API fits this feature" than "build a machine learning platform." This is a genuine advantage for application teams whose main job is shipping product features that happen to call an LLM, rather than teams running their own model training pipelines.

Vertex AI feels more like a full MLOps platform that generative AI was added to, rather than a generative-AI product with MLOps bolted on. There are more concepts to learn up front — Workbench notebooks, Pipelines, the Model Registry — but that depth pays off for teams doing classical machine learning (fraud detection models, forecasting, recommendation systems) alongside their generative AI work, because both live in the same platform with shared monitoring and governance.

If your team's workload is entirely "call an LLM from an app," Bedrock's narrower scope is usually faster to get productive in. If your team already runs custom-trained models in production and wants generative AI to plug into the same MLOps discipline, Vertex's breadth is a real, not cosmetic, advantage.

Compliance and Enterprise Governance

Both platforms are built to pass enterprise security review, and both inherit their compliance posture from their parent cloud rather than certifying the AI layer separately.

Bedrock is HIPAA-eligible under an AWS Business Associate Agreement, and supports SOC 2 and FedRAMP High authorization, letting it serve regulated workloads including healthcare and federal government use cases, per AWS's compliance documentation.

Vertex AI inherits Google Cloud's SOC 2, ISO 27001, and HIPAA-eligible status; HIPAA workloads require a signed BAA, same as AWS, and Google's Standard and Plus platform editions add VPC Service Controls, customer-managed encryption keys, and data residency options aimed at FedRAMP High and data-sovereignty requirements, according to Google Cloud's compliance and security controls documentation.

Amazon Bedrock now hosts more than 60 large language models from 13 different providers across 42 AWS regions — making it less a single AI product and more a distribution channel that competing model labs pay to be inside, the same way retailers pay for shelf space at a supermarket. Source: [AWS Bedrock documentation](https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html).

Are Bedrock and Vertex AI both HIPAA compliant?

Both are HIPAA-eligible, meaning healthcare data can be processed on either platform once your organization signs a Business Associate Agreement (BAA) with AWS or Google Cloud respectively. Eligibility applies to specific in-scope services, not the entire platform automatically, so teams should confirm which Bedrock or Vertex features are covered under their signed BAA before sending protected health information to them.

Where Azure AI Foundry Fits Into the Picture

No comparison of enterprise AI clouds is complete without the third major player. Azure AI Foundry (formerly Azure AI Studio) is Microsoft's equivalent platform, and it holds one card neither Bedrock nor Vertex can match: exclusive enterprise cloud access to OpenAI's models, including the GPT-5 family and OpenAI's reasoning-focused "o-series" models. Azure AI Foundry also hosts many of the same open-weight models available elsewhere (Llama, Mistral, Phi) and integrates deeply with Microsoft 365, Power Platform, and Azure's own data services.

If your organization is Microsoft-centric and specifically wants OpenAI's models inside an enterprise-governed environment, Azure is the only path — that is a hard technical constraint, not a preference. If OpenAI's specific models are not a requirement, the decision reduces back to the Bedrock-versus-Vertex question this article covers.

Where People Get This Comparison Wrong

The most common mistake is treating this as a model-quality comparison — "which chatbot answers better" — when for enterprise buyers it is almost always an infrastructure and organizational-fit comparison instead. Model quality differences between top-tier options change every few months as new versions ship; your data location, existing cloud contracts, and compliance requirements do not change nearly as often.

A second common mistake is assuming pricing loyalty locks you in. In practice, both AWS and Google have cut headline prices multiple times within a single year, sometimes within weeks of each other, which means a cost comparison done in January can be stale by June. Budget for pricing to move, and revisit vendor comparisons at renewal time rather than assuming the numbers you signed with are permanent.

A third mistake is underestimating switching costs on the agent and RAG layers specifically. Base LLM API calls are relatively portable — most SDKs abstract the request format closely enough that swapping a model is a day's work. Knowledge Bases, Vector Search indexes, and agent policy definitions (Cedar policies on AgentCore, or Google's agent configuration) are not portable at all. If your organization builds a serious RAG or agent system, you are choosing a platform for that system's lifetime, not just for the current model generation.

Multi-Cloud Strategies: When Using Both Makes Sense

A growing number of enterprises don't pick one platform exclusively — they run Bedrock for some workloads and Vertex AI for others, deliberately. A common pattern: customer-facing applications that need Claude's reasoning quality run on Bedrock, while internal analytics tools that live inside BigQuery call Gemini directly through Vertex, because moving that analytical data out of BigQuery would cost more engineering time than maintaining two integrations.

The tradeoff is real: running both platforms means duplicated observability tooling, two separate billing and cost-monitoring setups, and two sets of engineers who need platform-specific expertise. For a large enterprise with genuinely separate teams and genuinely separate data stores, that overhead is often worth it. For a smaller engineering organization, it usually is not — the operational tax of maintaining two platforms tends to outweigh the benefit of picking the theoretically best model for each individual feature.

How to Decide

Use this as a starting framework rather than a rigid rule:

• You're on AWS, and you want Claude or the flexibility of many model providers in one place → Bedrock.
• You're on GCP, want Gemini specifically, or run heavy BigQuery analytics → Vertex AI.
• You're on Azure and specifically need OpenAI's models → Azure AI Foundry (neither Bedrock nor Vertex is an option here).
• You run classical machine learning pipelines alongside generative AI → Vertex AI's MLOps depth is a genuine structural advantage.
• You need agent governance with explicit, auditable permission policies → Bedrock AgentCore's Cedar-based policy model is purpose-built for that.
• You're multi-cloud or an independent startup with no existing cloud commitment → Pick whichever cloud already holds your data; the AI feature differences are rarely large enough to justify migrating a data lake just to reach a marginally better model API.

For most enterprises, the honest rule is this: the AI platform follows the data, not the other way around. Companies rarely move petabytes of existing data to chase a slightly cheaper token price or a marginally higher benchmark score, and both AWS and Google know that — it's why so much of their 2026 roadmap has focused on deepening integration with their own existing data services (S3 and BigQuery) rather than purely on chasing model benchmarks.

The takeaway

Bedrock and Vertex AI have converged on nearly identical feature checklists — model marketplaces, managed RAG, agent frameworks, and enterprise compliance are now table stakes on both platforms, and their prices move within roughly 10% of each other for comparable open models. The decision that actually matters is rarely which platform has the smarter model this quarter; it's where your organization's data, compliance obligations, and existing engineering skills already live. Pick the platform that shortens the distance between your data and your model, not the one that wins this month's benchmark leaderboard, because that leaderboard will look different again in six months.

---

FAQs

Q: Can a company use both Bedrock and Vertex AI at the same time? A: Yes. Many enterprises run Bedrock for some workloads, such as Claude-powered customer-facing features, and Vertex AI for others, such as Gemini calls embedded directly in BigQuery analytics. The tradeoff is duplicated infrastructure, split observability, and two teams needing platform-specific expertise.

Q: Which platform gives access to Claude? A: Both do. Amazon has invested billions directly in Anthropic, and new Claude models have historically landed on Bedrock at or near the same time Anthropic ships them. Vertex AI added Claude through a separate partnership in 2024 and typically receives new models on a similar but not identical timeline.

Q: Is Gemini available on AWS Bedrock? A: No. Gemini is a Google first-party model family and is only accessible through Google's own platforms, either Vertex AI for enterprise use or Google AI Studio for individual developers.

Q: Are Bedrock and Vertex AI HIPAA compliant? A: Both are HIPAA-eligible under a signed Business Associate Agreement with AWS or Google Cloud respectively. Eligibility applies to specific in-scope services rather than the entire platform automatically, so confirm which features are covered before processing protected health information.

Q: Which platform is cheaper, Bedrock or Vertex AI? A: Neither is categorically cheaper. For identical open-weight models, token prices typically sit within about 10% of each other. Both companies cut prices multiple times through 2026 — Claude Sonnet pricing on Bedrock and Gemini Flash pricing on Vertex both dropped substantially within the same year — so total cost depends more on data egress, storage, and integration effort than on the headline token rate.

Q: Can I fine-tune models on Bedrock or Vertex AI? A: Yes on both, though the scope differs. Bedrock supports fine-tuning on selected hosted models, such as Claude Haiku and Llama variants. Vertex AI offers broader custom training capabilities through its full MLOps pipeline, including training models from scratch, not just fine-tuning existing ones.

Q: What is Amazon Nova, and how does it differ from Claude on Bedrock? A: Nova is Amazon's own first-party model family (Micro, Lite, Pro, and Premier for text, plus Canvas for images, Reel for video, and Sonic for speech), built and priced by Amazon itself, whereas Claude is built by Anthropic and hosted on Bedrock under a partnership. Nova models are generally cheaper and positioned for high-volume, cost-sensitive workloads; Claude is positioned for tasks needing stronger reasoning quality.

Q: Does Vertex AI support agent frameworks the way Bedrock AgentCore does? A: Yes, through what Google now calls the Gemini Enterprise Agent Platform, previously branded Vertex AI Agent Builder and then briefly Agentspace. It supports building and deploying multi-step, tool-using agents with access controls, similar in scope to Bedrock AgentCore, though the two use different underlying permission models — AgentCore uses AWS's Cedar policy language, while Google leans on its existing identity and access management layer.

Q: Why would a company choose Vertex AI over Bedrock if it's not already on Google Cloud? A: The main reasons are wanting Gemini specifically, needing tight BigQuery integration for SQL-native model calls, or running classical machine learning pipelines that benefit from Vertex's deeper MLOps tooling (Workbench, Pipelines, Model Registry) alongside generative AI, rather than treating generative AI as a separate system.

Q: Do Bedrock and Vertex AI support OpenAI's GPT models? A: No. GPT models are exclusive to Microsoft Azure AI Foundry due to Microsoft's investment in and partnership with OpenAI. Organizations that specifically require GPT-5-class models must use Azure regardless of their preference between Bedrock and Vertex AI for other workloads.

---

Sources

• AWS Bedrock documentation — "What is Amazon Bedrock"
• AWS Bedrock pricing page
• Amazon's investment disclosure on Anthropic
• Google Cloud — Gemini Enterprise / Agentspace product page
• Google AI for Developers — Gemini API pricing
• Google Cloud — Compliance and security controls documentation

Related reading: What Is Anthropic? AI Safety Company, What Is Cohere? Enterprise LLM Provider, What Is Mistral AI? Europe's LLM Lab