Short answer. Choose the platform of the cloud you already operate: Bedrock on AWS, Vertex AI on Google Cloud, Microsoft Foundry on Azure. If you are cloud-neutral, decide by model catalogue, procurement and the identity model your security team can support. A second cloud for a cheaper token rate rarely survives review.
All three are model catalogues wrapped in a cloud control plane. They sell access to first-party and partner models, add agents, evaluation and safety controls, and bill through the cloud commitment you already have. The differences that matter are operational, not conversational.
Two names changed in 2026: Azure AI Foundry is now Microsoft Foundry, and Google presents Vertex AI as the Gemini Enterprise Agent Platform. This page keeps the familiar names used in URLs and buying conversations.
How we evaluated this
We weighted five criteria for an enterprise platform decision that will outlive a single model generation.
- Identity and network fit, 30%. Whether the platform uses the IAM, private connectivity and policy model the buyer already operates.
- Model access, 25%. Breadth of first-party and partner models, regional availability and the ability to pin a version.
- Agent and evaluation tooling, 20%. Managed agents, tool use, tracing, datasets and graders.
- Price transparency, 15%. Published token rates, batch and provisioned options, and regional price differences.
- Exit cost, 10%. How much identity, networking, agent and evaluation configuration is platform-specific.
We read official pricing and product documentation on 20 November 2026. We did not run latency tests, quality evaluations or a migration. We did not verify enterprise discounts, committed-use rates or every regional quota. No 2027 result is implied.
The comparison that decides the platform
| Dimension | AWS Bedrock | Google Vertex AI | Microsoft Foundry |
|---|---|---|---|
| Identity | AWS IAM | Google Cloud IAM | Microsoft Entra ID and Azure RBAC |
| Private networking | VPC endpoints and PrivateLink | VPC Service Controls | Private Link and VNet isolation |
| Model catalogue | First-party and partner models, including Anthropic, OpenAI, xAI and Moonshot | Gemini plus partner and open models | Microsoft says more than 10,000 models from Microsoft, OpenAI, Anthropic and others |
| Agents | Bedrock Agents | Agent development and managed runtime | Foundry Agent Service |
| Evaluation | Bedrock evaluations | Offline and online agent evaluation | Tracing and built-in evaluations |
| Governance | Guardrails | IAM, policies and content security | Entra, Azure Policy and content filters |
| Billing | Per token, batch, provisioned | Per token, batch or flex, provisioned | Per token, batch, provisioned PTUs |
| Best fit | AWS-native teams | Google Cloud-native teams | Microsoft-native teams |
Confirm the current catalogue, region and quota before committing.
Pricing: three comparable rates, one dated trap
To keep the comparison honest, we used one mainstream text model per platform at its published global standard rate. These are list prices per million tokens, not negotiated rates.
| Platform and model | Input / 1M | Output / 1M | Context note |
|---|---|---|---|
| Bedrock: Claude Sonnet 5.5 | $2.00 | $10.00 | Global cross-region inference, US East (Ohio) |
| Vertex AI: Gemini 4 Carbon | $2.00 | $10.00 | Through 30 November 2026; $4.00 / $20.00 from 1 December 2026 |
| Foundry: GPT-6.1 Sol | $2.00 | $10.00 | Global short-context deployment; long context is $4.00 / $15.00 |
Assume 40 million input and 10 million output tokens a month, a 4:1 mix, list prices, no batch discount, no prompt caching and no committed use. At the current rates, each platform costs $180 a month: $80 for input and $100 for output. At 200 million input and 50 million output, the same math gives $900.
Gemini 4 Carbon breaks the tie from 1 December 2026. At $4.00 input and $20.00 output, the small workload becomes $360 and the large one $1,800. A 2027 budget that uses the $2/$10 rate would understate Google's list cost by half. Bedrock's Claude Sonnet 5.5 batch pricing was listed as not applicable in the region we checked, and Azure's GPT-6.1 Sol long-context band costs more, so model choice and context length move the bill more than the headline rate.
Identity and networking decide more than tokens
If your workloads already run in one cloud, the platform inherits an identity model your team already audits. Bedrock uses IAM policies and VPC endpoints. Vertex AI uses Google Cloud IAM and VPC Service Controls. Foundry uses Entra ID, Azure RBAC, Azure Policy and private networking. Rebuilding that in a second cloud is a security project, not a procurement line.
Private connectivity is the same story. The question is not whether a platform supports private endpoints, but whether your network team can operate the specific construct without a new exception. That is why a cheaper token on the wrong cloud often loses.
Agents and evaluation: convergence, not differentiation
All three now ship managed agents, tool use, tracing and evaluation. Bedrock has Agents and evaluations; Google documents agent development, a managed runtime, RAG Engine and offline or online evaluation; Foundry has Agent Service, a toolbox and built-in observability. The gap is smaller than the marketing suggests.
The real difference is where the agent's identity and data live. An agent that must act on an AWS resource is simpler on Bedrock. One that must respect Google Cloud policies is simpler on Vertex. One that must inherit Entra groups is simpler on Foundry. Portability of agent code is improving, but permissions are still platform-shaped.
Data residency and exit cost
All three publish regional deployment options and global endpoints, and all three price non-global endpoints at a premium. Bedrock distinguishes global and geo or in-region inference. Vertex prices global and non-global endpoints separately. Foundry offers global, data-zone and regional deployment. The names differ; the trade-off is the same: global routing is cheaper, regional processing is easier to defend.
Exit cost is highest in the layers around the model. The chat call is a base URL and a model ID. Identity, private networking, guardrails, agent tool schemas and evaluation datasets are not. Plan a parallel run and re-test tool calling and JSON mode before moving production traffic.
When each platform is the better choice
Bedrock is the better choice when AWS is your operating cloud, IAM and PrivateLink are mandatory, and you want Anthropic, OpenAI, xAI, Moonshot or open-weight models behind AWS controls. Vertex AI is the better choice when Google Cloud is the operating cloud, Gemini is central to the product, or you need Google's data and search stack close to the model. Foundry is the better choice when Microsoft is the operating cloud, Entra ID and Azure Policy are non-negotiable, or your applications already live in the Microsoft estate.
If you are genuinely cloud-neutral, choose by the model you need and the contract you can sign. Do not choose all three. Multi-cloud model access is usually better served by one OpenAI-compatible endpoint than by three control planes.
Where Plugsky fits
Plugsky is not a hyperscaler platform. It does not replace IAM, VPC Service Controls or Entra ID. It replaces the model layer with one OpenAI-compatible API, 39 public model IDs, a 262,144-token context on the main text models, and flat monthly plans. Deployment can be Plugsky cloud, a private VPC, on-premises or air-gapped, and the live regions are Frankfurt, the UAE, Singapore and Virginia, with Riyadh on Enterprise. See the model catalogue and sovereign deployment options.
What we did not verify
We did not verify throughput, uptime, model quality or agent reliability. We did not test a migration, and no statement here is a benchmark. We did not verify enterprise discounts, committed-use pricing, marketplace rates, taxes or every regional quota. We did not confirm that every catalogue model is available in every region, or that every agent feature is generally available rather than in preview.
For Plugsky, we did not verify a SOC 2 certification. A SOC 2 Type II report is available under NDA; confirm the current scope and audit period. The 99.9% figure is an uptime target, and the contract schedule governs. Private deployment is contract-led; confirm scope in writing.
Frequently asked questions
Which is better: Bedrock, Vertex AI or Microsoft Foundry?
Choose the platform of your operating cloud. Bedrock fits AWS, Vertex AI fits Google Cloud and Foundry fits Azure. If you are cloud-neutral, choose by model catalogue, procurement and identity model.
Do all three support multiple model providers?
Yes, to different degrees. Bedrock, Vertex AI and Foundry each publish a catalogue of first-party and partner models. The exact list changes, so check the current catalogue before designing around a model.
Which platform is cheapest?
No universal answer. At the rates we checked, Claude Sonnet 5.5 on Bedrock, Gemini 4 Carbon on Vertex and GPT-6.1 Sol on Foundry all cost $2 input and $10 output per million tokens. Gemini 4 Carbon rises to $4 and $20 on 1 December 2026.
How hard is migration between the three?
The model call is easy. Auth, endpoints, model IDs, safety controls, evaluation datasets and agent tooling are the hard parts. Plan a parallel run and re-test tool calling and JSON mode.
Is there a cloud-neutral alternative?
Yes. An OpenAI-compatible API such as Plugsky gives one endpoint across models and can run on its cloud, a private VPC, on-premises or air-gapped. The trade-off is leaving cloud-native procurement and IAM integration.
Does Azure AI Foundry still exist under that name?
Microsoft now brands the platform Microsoft Foundry. The Azure AI Foundry name and URL remain in use, and existing Azure OpenAI resources can be upgraded to a Foundry resource.
Sources
- Amazon Bedrock pricing — checked 20 November 2026.
- Google Cloud generative AI pricing — checked 20 November 2026.
- Azure OpenAI pricing — checked 20 November 2026.
- Microsoft Foundry overview — checked 20 November 2026.
- Google Cloud agent platform documentation — checked 20 November 2026.
Last checked 20 November 2026. Verify prices, regions and quotas before committing. Not paid.