Every major cloud vendor now sells the same basic pitch: bring your data, pick a model, ship an AI product without managing GPUs yourself. Amazon Bedrock, Microsoft’s Azure AI Foundry, and Google’s Vertex AI all promise that. But the three platforms diverge hard once you look past the marketing page. One stocks 1,700-plus models and locks up exclusive access to OpenAI’s GPT-5 family. Another ships with roughly a tenth of that catalog but built the deepest compliance paper trail of the three. The third gives you a free tier and the fastest custom-training pipeline, at the cost of not hosting OpenAI’s frontier models at all.
This comparison breaks down pricing, model catalogs, agent tooling, compliance, and real deployment data across all three platforms, using published vendor pricing, third-party cost analyses, and named customer case studies. If you’re choosing a home for a production AI workload in 2026, here’s what actually separates Amazon Bedrock, Azure AI Foundry, and Google Vertex AI.
None of this is a purely academic exercise. Whichever platform wins your workload also decides your default model roster, your agent framework, and how your security team documents compliance for the next few years. Getting the pick wrong doesn’t just cost money, it costs a rebuild. That’s the case for reading past the pricing page before committing engineering time to any one of the three.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Are Amazon Bedrock, Azure AI Foundry, and Vertex AI?
All three products solve the same basic problem: give enterprise teams one managed API surface for foundation models, instead of forcing them to run inference infrastructure themselves. But they grew out of different starting points, and that history still shapes how each one behaves today.
Amazon Bedrock
Amazon Bedrock is AWS’s fully managed generative AI service, reachable through a single API regardless of which model provider sits behind it. It launched into general availability in 2023 and has since become AWS’s answer to the “which model do I pick” problem, hosting Anthropic’s Claude, Meta’s Llama, Mistral, Cohere, DeepSeek, and Amazon’s own Titan and Nova models side by side. Its biggest structural advantage is depth of AWS integration. IAM roles, VPC networking, CloudTrail logging, and KMS encryption all wire in the way they would for any other AWS service, because Bedrock is built as one rather than bolted on afterward.
That AWS-native design shows up most clearly in how security teams evaluate Bedrock. A workload that already passes an AWS security review rarely needs a separate audit just because it now calls a foundation model, since the same logging, encryption, and network isolation controls already apply. That’s a smaller detail than model quality, but it’s often the reason a platform decision gets made in weeks instead of quarters.
Azure AI Foundry
Azure AI Foundry is Microsoft’s rebrand and expansion of what used to be Azure AI Studio and Azure OpenAI Service, rolled into a single portal and SDK surface. Its defining feature is exclusivity. Under Microsoft’s commercial partnership with OpenAI, Azure AI Foundry is the only major cloud platform offering first-party access to the full GPT-5 family, including GPT-5, GPT-5 mini, GPT-5 nano, GPT-5.3, GPT-5.4, and GPT-5.5. Foundry also carries the largest overall model catalog of the three, and it plugs directly into Microsoft 365, Teams, SharePoint, and Copilot, which matters most to organizations already standardized on Microsoft’s productivity stack.
Microsoft’s own Azure AI Foundry documentation frames the platform less as a model API and more as a full production lifecycle tool, covering evaluation, red-teaming, and observability alongside inference. That framing lines up with who actually buys it: large enterprises that need governance and audit trails around AI usage as much as they need the models themselves.
Google Vertex AI
Vertex AI is Google Cloud’s unified machine learning platform, home to the Gemini model family and Google’s open-weight Gemma line, plus a curated Model Garden that also carries Claude, Llama, and Mistral. Vertex leans harder into the full ML lifecycle than the other two, with strong AutoML tooling, tight integration with BigQuery for teams already doing analytics on Google Cloud, and the distinction of being the platform that originated the Agent-to-Agent (A2A) protocol, a Google-led open standard for letting AI agents from different vendors talk to each other.
Vertex AI also carries the most academic pedigree of the three. Google’s research organization, the same group behind the original Transformer architecture that every one of these platforms ultimately depends on, feeds directly into Vertex’s tooling roadmap. That heritage is part of why Vertex tends to win the argument for teams whose primary job is training and evaluating models, not just calling a hosted chat endpoint.
Full Spec Comparison: Bedrock vs Azure AI Foundry vs Vertex AI at a Glance
Before diving into pricing and benchmarks, here’s how the three platforms stack up across the specs that actually decide a platform choice.
| Spec | Amazon Bedrock | Azure AI Foundry | Google Vertex AI |
|---|---|---|---|
| Parent cloud | AWS | Microsoft Azure | Google Cloud |
| Model catalog size | 100+ models | 1,700+ models | 200+ curated models |
| Exclusive flagship access | Amazon Titan, Amazon Nova | OpenAI GPT-5 family (first-party) | Gemini 2.5 Pro, Gemma 4 |
| Pricing unit | Per-token, serverless | Per-token or Provisioned Throughput Units (PTUs) | Per-token, per-character (select tools), or free tier |
| Free tier | No | No (limited trial credits only) | Yes, for development and low-volume use |
| Fine-tuning support | Yes | Yes | Yes, plus modular billing per lifecycle stage |
| Agent framework | Bedrock Agents + AgentCore + Strands SDK | Azure AI Agent Service + Semantic Kernel | Agent Builder + Agent Development Kit (ADK) + Agent Engine |
| Agent runtime billing | $0.0895 per vCPU-hour (AgentCore) | No separate runtime fee, pay for inference and tool compute | ADK is free, pay for underlying Cloud Run/GKE infrastructure |
| Cross-vendor agent protocol | Supports A2A as a participant | Supports A2A as a participant | Originated the A2A protocol (April 2025) |
| Deepest compliance documentation | SOC 1/2/3, HIPAA-eligible, FedRAMP High (GovCloud), ISO 27001, GDPR | Inherits Azure’s broader compliance program | Inherits Google Cloud’s broader compliance program |
| Native data warehouse tie-in | Amazon Redshift, S3 | Microsoft Fabric, OneLake | BigQuery (tightest integration of the three) |
| Strongest enterprise ecosystem tie-in | Existing AWS accounts and IAM | Microsoft 365, Teams, Copilot, Sentinel | Google Workspace, BigQuery, Looker |
| Best documented for regulated industries | Finance (NYSE, Robinhood) | Manufacturing, retail (BMW, Coca-Cola) | Banking, retail (Wells Fargo, Target) |
The pattern that jumps out immediately: Azure AI Foundry wins on raw catalog size and OpenAI exclusivity, Amazon Bedrock wins on published compliance granularity and serverless pricing simplicity, and Vertex AI wins on data-pipeline integration and the only real free tier among the three.
Model Catalogs Compared: Which Platform Has Your Model?
Catalog size is the first thing most teams check, and the gap here is the widest of any metric in this comparison. Azure AI Foundry lists more than 1,700 models, by far the largest number of the three, spanning OpenAI’s GPT-5 line, Meta’s Llama, Mistral, Microsoft’s own Phi, Alibaba’s Qwen, DeepSeek, Cohere, and Anthropic’s Claude. Google Vertex AI’s Model Garden runs smaller and more curated, at roughly 200-plus models, anchored by the Gemini family and the newer Gemma 4 open-weight line, alongside Claude, Llama, and Mistral. Amazon Bedrock sits at the other end with just over 100 models, but that number understates its relevance, since it covers the models most production teams actually reach for: Claude, Llama, Mistral, Cohere, DeepSeek, and Amazon’s own Titan and Nova families.
The more consequential split is exclusivity, not raw count. As of 2026, OpenAI’s frontier GPT-5 models are available as first-party offerings only through Azure AI Foundry and OpenAI’s own direct API. Neither Bedrock nor Vertex AI hosts GPT-5 natively. That single fact decides the platform choice for a meaningful share of enterprise buyers before any other comparison even starts. If your organization has committed to GPT-5 specifically, Azure AI Foundry is not just the best option, it is close to the only one. Everyone else is choosing between the open-weight and third-party models that genuinely do overlap across all three: Llama, Mistral, and DeepSeek show up on every platform’s catalog, which caps how much lock-in you take on if you build around those models instead.
Pricing Breakdown: Per-Token Costs, Provisioned Throughput, and Hidden Fees
Sticker pricing on flagship models looks closer than most buyers expect. The bigger cost differences show up in the pricing unit itself, not the headline per-token rate.
| Pricing Dimension | Amazon Bedrock | Azure AI Foundry | Google Vertex AI |
|---|---|---|---|
| Flagship model example | Claude Sonnet 4.6 | GPT-5 | Gemini 2.5 Pro |
| Input price per million tokens | $3.00 | $1.25 | $1.25 (up to 200K context) |
| Output price per million tokens | $15.00 | $10.00 | $10.00 (doubles above 200K context) |
| Cheapest listed model option | Varies by open-weight model | GPT-5 nano tier | Gemma 4 26B at roughly $0.13 per million tokens |
| Billing model | Serverless per-token by default | Per-token or Provisioned Throughput Units | Per-token, per-character (select agent tools), or provisioned |
| Free tier | None | Trial credits only | Yes, for development and low-volume production |
| Agent runtime cost | $0.0895 per vCPU-hour (AgentCore) | No added runtime fee beyond inference/tools | ADK is free, infrastructure billed separately |
| Best fit workload for cost | Sustained, high-volume serverless inference | Predictable, steady-state enterprise traffic on PTUs | Bursty or experimental workloads using the free tier |
Notice that GPT-5 on Azure AI Foundry and Gemini 2.5 Pro on Vertex AI carry identical headline pricing, $1.25 per million input tokens and $10 per million output tokens, at least below the 200K context threshold where Vertex’s rate doubles. Claude Sonnet 4.6 on Bedrock runs higher on paper, at $3 and $15 respectively. But sticker price is not the full story. According to a third-party cost analysis of mid-volume workloads, Bedrock’s serverless pricing came out 15 to 25 percent cheaper than equivalent Azure or Vertex setups for teams processing 10 to 50 million tokens a month, largely because Bedrock’s model avoids the idle-capacity charges that provisioned throughput plans on Azure can carry during quiet periods.
Azure’s PTU model flips that math for teams with steady, predictable traffic. Paying for reserved throughput capacity gets cheaper than per-token billing once volume is high and consistent enough. Vertex AI’s published pricing page confirms the free tier is the outlier feature here. Neither Bedrock nor Azure AI Foundry offers anything comparable for teams that want to prototype before committing spend, which makes Vertex the default pick for early experimentation even among teams that expect to deploy elsewhere in production.
One more line item teams tend to miss during budgeting: none of the three platforms bundles data egress, embedding storage, or vector search indexing into the headline token price. AWS’s Bedrock pricing page breaks these out separately, and the other two follow the same pattern. Model a pilot project’s full cost, not just its inference cost, before comparing platforms on price alone. Teams that skip this step are usually the ones who end up surprised by a bill that runs 20 to 30 percent over their original estimate once storage and retrieval costs are added in.
FinOps and Cost Governance Across All Three Platforms
Token pricing gets the attention, but the bigger risk for most finance teams is runaway spend nobody notices until the monthly bill lands. Generative AI usage tends to spike unpredictably as adoption spreads inside an organization, and none of the three platforms enforces spending discipline by default. That’s a policy layer your team has to build on top, not a feature you get out of the box.
Practical guardrails look similar across all three platforms, even though the implementation details differ. Set hard budget alerts before granting broad access, tag every model call by team and project so cost allocation doesn’t turn into a forensic exercise later, and separate experimentation spend from production spend using different accounts, subscriptions, or projects depending on which cloud you’re on. Bedrock’s serverless-by-default pricing makes runaway cost easier to catch early, since there’s no idle reserved capacity racking up charges in the background. Azure’s PTU model is the opposite case: reserved throughput has to be actively monitored, because it bills whether or not it’s fully used. Vertex AI’s free tier is generous for prototyping, but teams routinely forget to set an upgrade trigger and get blindsided the month usage crosses into paid territory.
This is the same discipline that’s reshaping cloud budgeting more broadly. Our deeper look at how FinOps practices are helping CFOs tame runaway cloud costs covers the governance side of this problem in more detail, and the same principles apply directly to AI platform spend: visibility first, guardrails second, optimization third. Skipping straight to optimization without visibility is the most common mistake teams make when a generative AI pilot turns into a production rollout.
Performance Benchmarks: Latency, Throughput, and Cost-Efficiency
Hard, apples-to-apples latency benchmarks across all three platforms are scarce, since none of the three vendors publishes head-to-head numbers against competitors. What exists instead is a set of independently reported comparative claims, pulled from separate cost and performance analyses rather than a single controlled test.
- Cost-efficiency at mid-volume: Bedrock’s serverless architecture ran 15 to 25 percent cheaper than Azure or Vertex equivalents for workloads in the 10 to 50 million token per month range, according to a comparative cost analysis of the three platforms.
- Inference latency: Independent platform reviews put Bedrock’s Claude and Llama model latency at sub-200 millisecond response times for latency-sensitive production use, edging out the other two on raw response speed for those specific model families.
- Custom training speed: Vertex AI’s AutoML tooling was reported to cut training time by 40 to 60 percent compared to manual pipelines on the other platforms, reflecting Google’s investment in automated hyperparameter search and Google’s TPU hardware advantage for training jobs.
- Batch throughput: Vertex AI’s TPU-backed infrastructure showed a clear throughput edge for batch inference workloads exceeding 10,000 requests per hour, a scenario where Google’s custom silicon outperforms general-purpose GPU-backed inference.
- Provisioned throughput economics: Azure AI Foundry’s PTU pricing was flagged as expensive relative to serverless options on Bedrock and Vertex specifically for bursty, unpredictable traffic patterns, reversing to an advantage only once traffic volume is high and steady.
Read together, the pattern holds up across sources: Bedrock wins on cost and latency for steady serverless inference, Vertex wins on training speed and batch throughput thanks to TPU hardware, and Azure’s economics depend entirely on how predictable your traffic is. None of the three is faster or cheaper across every scenario, which is exactly why workload shape, not brand loyalty, should drive the decision.
Fine-Tuning and Model Customization
All three platforms support fine-tuning, and all three let you customize both commercial and open-weight models rather than restricting customization to Amazon’s, Microsoft’s, or Google’s own first-party models. Where they diverge is in how the billing and workflow are structured. Vertex AI takes the most explicit approach, breaking the machine learning lifecycle into separate billed stages: training, tuning, and inference are metered independently, which gives finance teams a clearer paper trail but also means costs can stack up in ways that are easy to underestimate during a proof of concept.
Bedrock and Azure AI Foundry both fold fine-tuning into a more unified pricing structure tied to the base model being customized, which is simpler to reason about but offers less visibility into which lifecycle stage is driving cost. For teams that plan to fine-tune extensively, this is worth testing early with a small representative dataset rather than committing a full production dataset to a platform before confirming the output quality justifies the spend.
Fine-tuned models and custom embeddings are not portable between platforms, a detail that matters more during the migration planning covered later in this piece than it does during initial platform selection. A model fine-tuned on Bedrock has to be retrained from the base weights if it later needs to run on Azure AI Foundry or Vertex AI, since none of the three vendors exports fine-tuning deltas in a format the others can read. Treat that as a switching cost baked into the decision from day one, not a problem to solve later.
Agent Building and Orchestration Frameworks
Agent tooling is where all three platforms have shipped the fastest over the past year, and it’s now a bigger differentiator than raw model access for teams building anything beyond a simple chatbot. Amazon Bedrock combines Bedrock Agents with the newer AgentCore runtime and the Strands SDK, giving developers the most model-flexible agent stack inside AWS and what several reviewers flagged as the strongest fit for zero-trust, multi-tenant deployments, thanks to vault-backed token management. AgentCore reached broad enterprise general availability between October 2025 and the first quarter of 2026.
Azure AI Foundry’s Agent Service, paired with Microsoft’s Semantic Kernel framework, reached enterprise general availability on a similar timeline and reports more than 10,000 Agent Service customers at GA. Its clearest advantage is depth of integration with Microsoft 365 data sources: agents built on Foundry can reach into Outlook, Teams, SharePoint, and Microsoft Sentinel in ways the other two platforms simply cannot replicate without custom connector work. Foundry also skips a separate runtime fee altogether, charging only for the underlying model inference and tool calls.
Google Vertex AI shipped its Agent Development Kit, or ADK, in the second quarter of 2026, and it’s been described as the cleanest agentic primitive surface of the three, having already logged more than 7 million downloads. Vertex’s bigger structural advantage is that Google originated the Agent-to-Agent (A2A) protocol back in April 2025, an open standard for letting agents built on different vendors’ platforms communicate, and Vertex was the first platform to ship a managed runtime built around it. Bedrock and Azure both support A2A as participants, but Vertex built it. Worth noting for context: Anthropic itself entered the managed agent space with its own public beta in April 2026, a sign that this category is still shifting fast enough that today’s leader is not guaranteed to hold that position by year’s end.
| Agent Attribute | Amazon Bedrock | Azure AI Foundry | Google Vertex AI |
|---|---|---|---|
| Core framework | Bedrock Agents + AgentCore | Azure AI Agent Service | Agent Builder + Agent Engine |
| Developer SDK | Strands SDK | Semantic Kernel | Agent Development Kit (ADK) |
| General availability | Oct 2025 – Q1 2026 | Oct 2025 – Q1 2026 | ADK shipped Q2 2026 |
| Strongest data source tie-in | AWS-native services | Outlook, Teams, SharePoint, Sentinel | BigQuery, Vertex AI Search |
| Cross-vendor protocol role | A2A participant | A2A participant | A2A originator (April 2025) |
| Reported adoption signal | Fastest YoY growth cited since GA | 10,000+ Agent Service customers at GA | 7M+ ADK downloads |
Compliance, Security, and Data Governance
Amazon publishes the most granular compliance documentation of the three platforms specifically for Bedrock. It’s in scope for AWS’s SOC 1, 2, and 3 reports, HIPAA-eligible under AWS’s Business Associate Agreement once a customer accepts the BAA and configures workloads to AWS guidance, FedRAMP High authorized in the AWS GovCloud (US-West) region, in scope for FedRAMP Moderate more broadly, certified under ISO/IEC 27001:2022, and covered automatically by AWS’s GDPR Data Processing Addendum.
| Certification | Amazon Bedrock | Azure AI Foundry | Google Vertex AI |
|---|---|---|---|
| SOC 2 | Confirmed in scope | Inherited from Azure | Inherited from Google Cloud |
| HIPAA eligibility | Confirmed, via BAA | Inherited from Azure | Inherited from Google Cloud |
| FedRAMP High | Confirmed, GovCloud US-West | Inherited from Azure Government | Inherited from Google Cloud |
| ISO/IEC 27001 | Confirmed (2022 revision) | Inherited from Azure | Inherited from Google Cloud |
| GDPR | Confirmed, automatic DPA | Inherited from Azure | Inherited from Google Cloud |
That doesn’t mean Azure and Google fall short on compliance in practice. Both fold their AI platforms into compliance programs that are, at the whole-cloud level, as mature as AWS’s. Microsoft’s Azure Government cloud and Google Cloud’s own FedRAMP High authorizations are well established. The distinction is documentation granularity at the AI-service layer specifically. Amazon has published a service-specific compliance breakdown for Bedrock, while equivalent AI-service-specific pages for Azure AI Foundry and Vertex AI were harder to independently confirm at the same level of detail during this research. If compliance documentation is a hard procurement requirement, ask your account team for the AI-service-specific paperwork rather than assuming the parent cloud’s certification automatically covers the AI layer identically.
Data residency is the other question every regulated buyer eventually asks: where does a prompt physically go once it leaves your application. All three vendors let customers pin processing to a chosen region and generally commit not to use customer prompts to train their own base models by default on enterprise tiers, though the exact contractual language differs by platform and by contract tier. Read the data processing addendum for your specific account type rather than assuming a blog post, including this one, covers the fine print for your exact contract.
Real-World Deployments: Who’s Using Which Platform
Vendor case studies always come with a promotional slant, but they’re still the clearest public signal of what each platform is actually used for in production, as opposed to what it’s marketed for. Here’s a sample of named deployments across all three, drawn from published AWS, Microsoft, and Google customer materials, including AWS’s own Bedrock customer directory.
| Company | Platform | Sector | Reported Use Case |
|---|---|---|---|
| Robinhood | Amazon Bedrock | Fintech | Scaled Bedrock usage from 500 million to 5 billion tokens a day in six months, per AWS’s customer materials |
| New York Stock Exchange | Amazon Bedrock | Finance | Uses Bedrock’s models to answer regulatory and compliance questions across billion-transaction daily operations |
| Adobe | Amazon Bedrock | Software | Reported a 20 percent improvement in developer support search accuracy |
| AstraZeneca | Amazon Bedrock | Pharmaceuticals | Uses Bedrock Agents to accelerate drug development research and data analysis |
| HappyFox | Amazon Bedrock | Customer support software | Reported a 40 percent efficiency gain in automated ticketing and a 30 percent lift in agent productivity |
| Adidas | Azure AI Foundry | Retail | Generates personalized marketing content and product descriptions for e-commerce |
| BMW | Azure AI Foundry | Automotive | Powers customer service chatbots and internal technician knowledge tools |
| Coca-Cola | Azure AI Foundry | Consumer goods | Analyzes social sentiment and generates marketing copy, plus supply chain support |
| Hilti | Azure AI Foundry | Manufacturing | Predictive maintenance for construction equipment using sensor data |
| Mitsubishi Electric | Azure AI Foundry | Manufacturing | Automated visual quality control on production lines |
| Wells Fargo | Google Vertex AI | Banking | Deploying AI agents company-wide, per Google Cloud’s published materials |
| Netflix | Google Vertex AI | Media | Content recommendations and dynamic ad targeting trained on user interaction data |
| Uber | Google Vertex AI | Transportation | Route optimization, demand forecasting, and driver-passenger matching |
| Target | Google Vertex AI | Retail | Inventory forecasting using historical sales and external demand signals |
| Siemens | Google Vertex AI | Industrial | Predictive maintenance on industrial equipment via sensor stream analysis |
A pattern worth calling out: Bedrock’s named customers skew toward regulated finance (Robinhood, NYSE) and enterprise software, Azure AI Foundry’s skew toward existing Microsoft-ecosystem manufacturers and consumer brands, and Vertex AI’s skew toward companies with data-heavy, analytics-native operations (Netflix, Uber, Target). That’s not a coincidence. Each platform’s customer base tends to mirror the cloud infrastructure those companies were already running before they added generative AI on top.
That gravitational pull toward your existing cloud is worth taking seriously as a decision input in its own right. Migrating a company’s core data infrastructure just to chase a marginally better AI platform is rarely worth the disruption. In nearly every case study above, the company picked the AI platform that matched the cloud it was already running, not the other way around. Treat your current infrastructure as the strongest prior in this decision unless you have a specific, well-documented reason to override it.
Getting Started: Code Examples for Each Platform
The SDK experience differs enough between platforms that it’s worth seeing side by side before committing engineering time. Below are minimal, illustrative examples for invoking a model on each platform using its native Python SDK. Treat these as a starting shape, not copy-paste production code, since exact method signatures shift between SDK versions.
Amazon Bedrock (boto3)
import boto3
import json
client = boto3.client("bedrock-runtime", region_name="us-east-1")
response = client.invoke_model(
modelId="anthropic.claude-sonnet-4-6",
body=json.dumps({
"anthropic_version": "bedrock-2023-05-31",
"max_tokens": 512,
"messages": [{"role": "user", "content": "Summarize this quarter's cloud spend."}]
})
)
result = json.loads(response["body"].read())
print(result)Azure AI Foundry
from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential
project = AIProjectClient(
endpoint="https://your-foundry-project.services.ai.azure.com",
credential=DefaultAzureCredential()
)
client = project.inference.get_chat_completions_client()
response = client.complete(
model="gpt-5",
messages=[{"role": "user", "content": "Summarize this quarter's cloud spend."}]
)
print(response.choices[0].message.content)Google Vertex AI
import vertexai
from vertexai.generative_models import GenerativeModel
vertexai.init(project="your-gcp-project", location="us-central1")
model = GenerativeModel("gemini-2.5-pro")
response = model.generate_content("Summarize this quarter's cloud spend.")
print(response.text)Notice the shape of each call reflects the platform’s philosophy. Bedrock’s raw boto3 call is the most AWS-native and verbose, matching its deep IAM and service-mesh integration. Azure AI Foundry leans on its project and identity abstractions, matching its enterprise-identity focus. Vertex AI’s call is the shortest of the three, consistent with Google’s push toward a simpler, more Pythonic developer experience.
Language support beyond Python also varies more than most teams expect going in. All three ship first-class SDKs for Python, JavaScript or TypeScript, and Java, which covers the large majority of production backends. Bedrock’s SDK surface is really the wider AWS SDK family, so any language AWS already supports, including Go, Ruby, and .NET, gets Bedrock access essentially for free. Azure AI Foundry mirrors that pattern through the broader Azure SDK ecosystem. Vertex AI’s official coverage is narrower outside Python, which matters if your production stack is built in something like Go or Rust and you were hoping to skip a REST wrapper.
Migration Guide: Moving Between Bedrock, Azure AI Foundry, and Vertex AI
Moving a production AI workload between these platforms is rarely a weekend project, mostly because agent orchestration and fine-tuned models don’t travel with you. Here’s a realistic sequence for planning the move.
- Audit current model dependencies. List every model your application calls directly, including any fine-tuned or custom-trained variants, and flag which ones exist on your target platform’s catalog versus which ones don’t transfer at all.
- Abstract the application layer. Put a thin routing layer between your application code and the model API, so swapping the underlying platform later doesn’t mean rewriting every call site. This step is cheap early and expensive to retrofit later.
- Re-plan fine-tuned models and embeddings separately. Fine-tuned weights and embedding indexes are not portable across Bedrock, Azure AI Foundry, and Vertex AI. Budget time to retrain or re-embed on the destination platform rather than assuming an export/import path exists.
- Rebuild agent orchestration logic from scratch. Bedrock Agents, Azure AI Agent Service, and Vertex Agent Builder are not compatible with each other. If your application relies on agent workflows, this is usually the single biggest line item in a migration budget, not the model calls themselves.
- Remap identity and security controls. IAM roles, IAM policies, and VPC configurations on AWS don’t translate to Azure’s identity model or Google Cloud’s IAM. Rebuild access controls natively on the destination platform instead of trying to mirror the source platform’s structure.
- Re-model costs against the new pricing unit. A workload priced against Bedrock’s serverless per-token model needs a fresh cost model if it’s moving to Azure’s PTU structure or Vertex’s mixed token/character billing. The unit economics rarely translate one to one.
- Run both platforms in parallel before cutover. Route a small percentage of production traffic to the new platform first, compare output quality and latency against the incumbent, and only fully cut over once that comparison holds up across a real traffic sample, not just a test set.
Most teams underestimate step four and overestimate step one. Model access is rarely the blocker. Rebuilding agent logic and re-testing fine-tuned model behavior on new infrastructure is where migration timelines actually slip.
Pros and Cons of Each Platform
Amazon Bedrock
Pros: broadest practical model marketplace under a single AWS bill, serverless pricing that avoids idle-capacity charges, the deepest published compliance documentation of the three, fastest reported adoption growth since GA, and the tightest fit for teams already standardized on AWS IAM and networking.
Cons: no first-party access to OpenAI’s GPT-5 family, agent tooling reportedly shipped slower than Vertex AI’s, AgentCore’s per-vCPU-hour billing adds a layer of cost complexity, and its data-analytics integration doesn’t match Vertex AI’s tie-in with BigQuery.
Azure AI Foundry
Pros: the largest model catalog by a wide margin, exclusive first-party access to GPT-5 and the rest of OpenAI’s frontier lineup, the deepest integration with Microsoft 365, Teams, and Copilot, and no separate agent runtime fee on top of inference costs.
Cons: Provisioned Throughput Unit pricing gets expensive fast for bursty or unpredictable traffic, the platform’s biggest advantages mostly apply to organizations already committed to the Microsoft ecosystem, and a catalog of 1,700-plus models includes plenty of long-tail options most teams will never touch.
Google Vertex AI
Pros: the only genuine free tier among the three for development and low-volume production, the fastest reported custom-training pipeline via AutoML, the strongest batch-inference throughput thanks to TPU hardware, and first-mover advantage on the A2A cross-vendor agent protocol.
Cons: no first-party OpenAI model access, a mid-sized catalog that trails Azure AI Foundry by a wide margin, and a TPU-first architecture that can mean less flexibility for teams whose tooling is built around GPU-based inference.
Which Platform Should You Choose? 7 Use-Case Recommendations
| If you are… | Choose | Because |
|---|---|---|
| A startup wanting the broadest model choice under one bill | Amazon Bedrock | 100+ models, serverless pricing, no minimum commitment |
| An enterprise already standardized on Microsoft 365 and Teams | Azure AI Foundry | Native Copilot, SharePoint, and Sentinel integration |
| A team whose stack is already built on BigQuery and Google Cloud | Google Vertex AI | Tightest data-to-model pipeline of the three |
| A regulated bank or fintech needing granular compliance paperwork | Amazon Bedrock | Published FedRAMP High, HIPAA, and ISO 27001 documentation at the service level |
| Locked into OpenAI’s GPT-5 family specifically | Azure AI Foundry | The only platform with first-party GPT-5 access |
| A research lab doing heavy custom model training | Google Vertex AI | AutoML speed and TPU-backed training throughput |
| A multi-cloud shop trying to avoid single-vendor lock-in | A hybrid of Bedrock and Vertex AI | Both support the shared A2A protocol and overlapping open-weight models like Llama and Mistral |
Cloud Market Share and Where This Is Heading in 2026
The AI platform layer doesn’t exist in a vacuum. It rides on top of each vendor’s broader cloud infrastructure business, and that underlying market is still consolidating around the same three names. According to Synergy Research Group, AWS held 28 percent of global cloud infrastructure spend in the most recent reported quarter, with Microsoft Azure at 21 percent and Google Cloud at 14 percent. Together, the three control 63 percent of the global market, a figure that lines up with Statista’s independent tracking of the same segment. AWS’s share has actually been sliding, down from 30 percent in the fourth quarter of 2024 to 28 percent in the most recent data, while Azure and Google Cloud have each picked up a point or two of share over the same stretch.
That shift matters for AI platform buyers specifically, because model access and infrastructure spend increasingly move together. As enterprises route more spend toward Azure and Google Cloud relative to AWS, expect Azure AI Foundry and Vertex AI to keep closing the adoption gap with Bedrock, even as Bedrock currently reports the fastest raw growth rate of the three since its 2023 launch. Our sister comparison of AWS vs Azure’s broader market share and archive storage costs covers the infrastructure layer underneath this decision in more depth, and it’s worth reading alongside this one if you’re negotiating a multi-year cloud commitment rather than just picking an AI API.
Pulling every category covered above into one scorecard makes the trade-offs easier to hold in your head at once.
| Category | Winner | Runner-Up |
|---|---|---|
| Model catalog size | Azure AI Foundry (1,700+) | Google Vertex AI (200+) |
| Exclusive frontier model access | Azure AI Foundry (GPT-5 family) | Google Vertex AI (Gemini, Gemma) |
| Cost at mid-volume, serverless | Amazon Bedrock (15-25% cheaper) | Google Vertex AI |
| Free tier for prototyping | Google Vertex AI | None |
| Published compliance granularity | Amazon Bedrock | Azure AI Foundry |
| Custom training and AutoML speed | Google Vertex AI (40-60% faster) | Amazon Bedrock |
| Agent protocol leadership | Google Vertex AI (originated A2A) | Amazon Bedrock (AgentCore) |
| Enterprise productivity suite integration | Azure AI Foundry (Microsoft 365) | Google Vertex AI (Workspace) |
| Reported adoption growth rate | Amazon Bedrock (180% YoY cited) | Azure AI Foundry |
No single platform sweeps the scorecard, which is the whole point. The right pick tracks your existing cloud footprint and your specific model requirements far more than it tracks any single vendor’s marketing claims.
The Verdict: Our Recommendation Backed by the Data
“It depends on your use case” is technically true and mostly useless, so here’s a real recommendation. For a typical enterprise team in 2026 that isn’t already locked into one cloud, Amazon Bedrock is the strongest default starting point. It offers the broadest practical model selection, the most granular compliance documentation, serverless pricing that’s 15 to 25 percent cheaper than the alternatives at mid-volume, and the fastest-growing customer base of the three platforms.
That default breaks in two directions. If your organization runs on Microsoft 365 and needs OpenAI’s GPT-5 family specifically, Azure AI Foundry isn’t just competitive, it’s close to mandatory, since it holds exclusive first-party access to those models and the deepest Copilot integration. If your data already lives in BigQuery and your team’s core need is fast custom model training rather than access to a specific chat model, Vertex AI’s AutoML speed and free tier make it the more rational choice, especially for a research-heavy or ML-native organization.
The one conclusion that holds regardless of which platform wins your workload: build an abstraction layer between your application and whichever model API you pick. Given how fast agent frameworks, pricing units, and model catalogs are all still moving across all three platforms, the team that avoids hard-coding itself to one vendor’s SDK today will have a much easier 2027.
One last data point worth weighing before you sign a multi-year commitment: none of these three platforms is standing still long enough for today’s comparison to be the last word. Vertex AI’s agent tooling shipped a full quarter after Bedrock’s and Azure’s, then immediately took the lead on cross-vendor protocol support. Azure AI Foundry’s model catalog grew past 1,700 entries in a matter of months. Bedrock added Nova and expanded its AgentCore runtime inside the same window. Revisit this comparison against your own workload roughly every two quarters, not once a year, because the gaps documented here are closing and reopening faster than most procurement cycles move.
Frequently Asked Questions
What is the main difference between Amazon Bedrock, Azure AI Foundry, and Vertex AI?
Bedrock focuses on model breadth and AWS-native compliance depth, Azure AI Foundry focuses on catalog size and exclusive OpenAI GPT-5 access, and Vertex AI focuses on data-pipeline integration and fast custom model training. All three support the major open-weight models like Llama and Mistral.
Can I use Claude, GPT, and Gemini on all three platforms?
Not quite. Claude is available on Bedrock and through Vertex AI’s Model Garden. GPT-5 is only available as a first-party model on Azure AI Foundry and OpenAI’s own API. Gemini is exclusive to Google Vertex AI. Open-weight models like Llama, Mistral, and DeepSeek are the ones that genuinely overlap across all three.
Which platform is cheapest for a startup just getting started?
Google Vertex AI is the easiest entry point because it’s the only one of the three with a genuine free tier for development and low-volume production use. For sustained production workloads at moderate volume, Amazon Bedrock’s serverless pricing tends to come out cheaper according to third-party cost comparisons.
Is Azure AI Foundry the only way to access GPT-5?
Among the three major cloud platforms, yes. As of 2026, GPT-5 and the rest of OpenAI’s frontier lineup are available as first-party models only through Azure AI Foundry or directly through OpenAI’s own API, not through Amazon Bedrock or Google Vertex AI.
Which platform is best for regulated industries like banking and healthcare?
Amazon Bedrock currently publishes the most detailed service-specific compliance documentation, including FedRAMP High authorization in AWS GovCloud and a clear HIPAA Business Associate Agreement path, and it counts NYSE and Robinhood among its named finance customers. Azure and Vertex both inherit strong compliance programs from their parent clouds, but ask your account team for AI-service-specific documentation rather than assuming parity by default.
Do all three platforms support multi-agent orchestration?
Yes, though the maturity differs. Vertex AI originated the Agent-to-Agent (A2A) protocol and was first to ship a managed multi-agent runtime around it. Bedrock’s AgentCore and Azure’s AI Agent Service both reached general availability on a similar timeline, between October 2025 and the first quarter of 2026, and both support A2A as participants rather than as the protocol’s originator.
How hard is it to migrate from one platform to another?
Harder than most teams expect. Model access itself is rarely the bottleneck since open-weight models overlap across all three. The real cost is rebuilding agent orchestration logic and retraining fine-tuned models, since neither transfers between Bedrock, Azure AI Foundry, and Vertex AI.
Which platform has the largest model catalog in 2026?
Azure AI Foundry, with more than 1,700 models listed, well ahead of Vertex AI’s 200-plus curated Model Garden and Bedrock’s 100-plus. Catalog size isn’t the whole story though. Bedrock’s smaller list is weighted toward the models most production teams actually use.
Related Coverage
- AWS vs Azure 2026: 31% vs 24% Market Share and a 75% Archive Cost Gap
- AWS vs Azure vs Google Cloud 2026: The Definitive Cloud Platform Comparison
- Opus 4.8 vs GPT-5.5 vs Gemini 3.1: 8-Point SWE Gap [2026]
- OpenAI on AWS Bedrock: $38B Pact Ends Azure Lock-In [2026]
- FinOps in 2026: How CFOs Are Finally Taming Runaway Cloud Costs
- Snowflake vs Databricks vs BigQuery 2026: $6.25/TB Showdown
For more coverage of the cloud platforms powering this shift, see Tech Insider’s full cloud computing archive.
