Claude on AWS Bedrock, Vertex AI and Azure: Models, Prices, Quotas and Gaps
Claude on Bedrock, Vertex AI and Microsoft Foundry compared: models, regions, quotas, auth, feature gaps and Claude Code setup. Checked 17-10-2026.
Published 24-10-2026. Last verified 17-10-2026. We re-check this page every month.
Quick answer
Claude runs on Amazon Bedrock, Google's Vertex AI (Anthropic's docs now call it Agent Platform) and Microsoft Foundry, and the right one is usually the cloud your legal or infrastructure team has already approved. This guide is for an engineering lead or platform engineer who must put Claude inside an existing cloud contract and needs to know what breaks. The three things that matter most: regional endpoints cost 10% more than global ones on Bedrock and Vertex, the Message Batches API and Claude Managed Agents are listed as unsupported on all three, and Foundry pay-as-you-go quotas start at 40 requests per minute for current Opus models. We could not read per-model cloud prices on 17-10-2026, so we use Anthropic's own prices as the reference and say where the cloud numbers are missing.
How we checked the information in this article
We read the provider documentation and Anthropic's platform pages on 17-10-2026. We ran no deployments ourselves for this article, so every number below is vendor-reported and none is measured by us. Our method in three lines: we read the primary page, we write down what it says with the date, and we say so when a page did not load or does not contain the answer.
Our gate for this page: a model counts as available on a platform only if a page we could open on 17-10-2026 lists it for that platform, and we label it "access unclear" when the same page gates it behind a form or verification. The gate rejected three things. We left out the Claude Mythos models because the pages call them limited availability or invitation only. We did not call Opus 5.5, Sonnet 5.5 and Haiku 5.5 open on the Bedrock Messages API endpoint, because Anthropic's page lists them as "See Access" or verification required. And we dropped any price comparison for the same model across clouds, because the AWS and Microsoft price pages did not give per-model amounts and Google's page stops at older models.
Which Claude models run on Bedrock, Vertex AI and Azure?
On the dates above, Anthropic's platform pages list current Claude models on all three clouds, but not the same set and not with the same access. The table shows the models we think most teams actually pick, with Anthropic's own API price as the reference (vendor-reported, from the Anthropic pricing page, 17-10-2026).
| Model | Bedrock | Vertex AI | Foundry, Azure-hosted | Foundry, Anthropic-hosted | Anthropic API price per million tokens (input / output) |
|---|---|---|---|---|---|
| Fable 5.1 | Listed | Listed | No | Yes | $10 / $50 |
| Opus 5.5 | Listed, access unclear | Listed | Yes | Yes | $4 / $20 |
| Sonnet 5.5 | Listed, access unclear | Listed | Yes | Yes | $2 / $10 |
| Haiku 5.5 | Listed, access unclear | Listed | Yes | Yes | $0.10 / $0.50 for prompts up to 100,000 tokens |
| Haiku 4.5 | Yes | Listed | Yes | Yes | $1 / $5 |
Each row sits here for a reason: Opus 5.5 is the model Claude Code picks by default on Bedrock when you pin nothing, Sonnet 5.5 and Haiku 5.5 are the cheaper working models, Fable 5.1 is the most expensive one in the set, and Haiku 4.5 is the one model with open access on every platform we read.
Two details change what you can deploy. On Foundry, "Hosted on Azure" and "Hosted on Anthropic" are different deployments with different feature sets, and Fable 5.1 only exists in the Anthropic-hosted variant. On Bedrock there are now two integrations: the older InvokeModel and Converse API with versioned model IDs, documented for Opus 4.6 and earlier, and the newer Claude in Amazon Bedrock endpoint that uses the Anthropic Messages API shape. The newer page says Fable 5.1, Fable 5, Opus 4.8, Sonnet 5, Opus 4.7 and Haiku 4.5 are open to all Bedrock customers. Check the exact list in Anthropic's Claude in Amazon Bedrock documentation before you plan around one model, and see Anthropic's Vertex AI and Microsoft Foundry pages for their lists (all listed under Sources).
Not tested by us: whether a given model is enabled in your own account or region. AWS moved its own model catalogue to a "models at a glance" page that we did not open, and we only saw the navigation of Google's partner-model page. Use the console of your cloud as the final check.
How much does Claude cost on Bedrock, Vertex AI and Azure?
For the same model, the cloud price is not always the Anthropic price, and on 17-10-2026 we could not confirm the difference for current models. What the pages do say, all vendor-reported:
Bedrock and Vertex AI both state that regional or multi-region endpoints carry a 10% premium over global endpoints for Sonnet 4.5 and newer models, and that global endpoints have no premium. Models up to Opus 4.1 keep their old price structure. Vertex also lists a multi-region endpoint for the US and EU with the same 10% premium.
Foundry bills through the Azure Marketplace in Claude Consumption Units, metered hourly and invoiced monthly in arrears, with no prepaid balance. Anthropic's pricing page says a US Data Zone Standard deployment on Azure gets the same 1.1x multiplier as the Anthropic API's US-only inference option. Microsoft's billing page, last updated 15-08-2026, says the unit has a fixed price on the Azure Marketplace offer but does not print the amount, so we cannot quote a per-model Foundry price.
Google's pricing page loaded for us without a date, and it lists Opus 4.1 at $15 input and $75 output per million tokens and Sonnet 4 at $3 and $15, but not the current models. AWS's pricing page showed a prompt-caching note, a cache read price for Sonnet 5.5 reduced by 50% from 07-10-2026, and a batch inference price 50% below on-demand for select models, but the Claude price tables did not appear in what we could read.
Our advice: do not choose a cloud on price from a blog post, including this one. Take the model you need, open your own console or price page, and compute a month of your real token volume. Our Claude API pricing guide gives the Anthropic-side reference, and our prompt caching guide shows where most of the cost goes. If cost spikes are your worry, the skill finops anomaly intelligence links AWS cost spikes to engineering events for a root cause report, and context window tracker shows context usage as you work in OpenClaw.
Where does your data stay: regions and residency?
All three offer a global option and a residency option, and the residency option is where they differ. If a regulator or customer contract forces a location, treat this section as a checklist for your legal team. We are not lawyers and this is not legal advice. Our legal and compliance skills help structure the review, and the official terms of each provider decide.
Bedrock: global endpoints route dynamically across regions, and regional endpoints resolve to the single AWS region you specify. Anthropic's page lists regional availability across regions in Europe, Asia Pacific, the Middle East, Africa, Canada, South America and the US. Claude Fable 5.1 regional endpoints are currently available in us-east-1 only.
Vertex AI: global, multi-region (currently US and EU) and regional endpoints. Regional endpoints are required for single-region residency or provisioned throughput. A code comment on Anthropic's page says specific regional endpoints support Sonnet 4.6 and earlier, and newer models use global or multi-region endpoints. That is the point to verify first if you need a single EU region for a current model.
Foundry: Global Standard and US Data Zone Standard. The page we read lists no EU data zone. Anthropic's page says Sonnet 5.5 supports only Global Standard, while Microsoft's quota page, updated 22-09-2026, lists claude-sonnet-5-5 with a US Data Zone Standard deployment. The two pages disagree, so we would not promise US-only processing for Sonnet 5.5 on Foundry until your Azure portal shows the deployment type.
All three pages say Anthropic recommends keeping your own activity logs for at least a 30-day rolling window, and that data handling on each platform is governed by the cloud provider.
Which features are missing compared with the Anthropic API?
Several, and they decide whether your agent design works. On the pages we read, all vendor-reported:
Unsupported on all three: the Message Batches API, Claude Managed Agents (see our Managed Agents guide), server-side fallback (use the client-side pattern), and the Models, Admin and Compliance endpoints. Unsupported on Bedrock and Vertex, and on Foundry's Azure-hosted deployments: the Files API, URL sources for images and documents, Agent Skills, the MCP connector and programmatic tool calling. Server-side code execution is unsupported on those same deployments, and web search is supported on Vertex but not on Bedrock.
Supported on Bedrock and Vertex: prompt caching, extended thinking, tool use with the bash, computer use, memory and text editor tools, and citations. Structured outputs are listed as supported on the older Bedrock integration and on Vertex, but not on the newer Bedrock Messages API endpoint. We did not see a prompt caching statement for Foundry, so we mark it as not confirmed.
One nuance on batch. Anthropic lists the Message Batches API as unsupported on Bedrock, while AWS's pricing page says select Anthropic models get batch inference at a price 50% below on-demand. These are different features, so check whether your model is on the AWS batch list before you rule batch out.
Context windows: Anthropic's pages give a 1 million token window on all three clouds for the Fable models, Opus 5.5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5.5, Sonnet 5, Sonnet 4.6 and Haiku 5.5, and 200,000 tokens for older ones. Bedrock limits request payloads to 20 MB and Vertex to 30 MB. For caching details see the prompt caching guide linked in the pricing section, and if your design leans on Agent Skills or the MCP connector, plan to keep that logic in your own code. Our API and backend skills cover the code side, and agent reliability audit turns agent traces and tool logs into a production-readiness report.
What are the quotas and rate limits, and how do you raise them?
Bedrock: for the Messages API endpoint, Anthropic's page states a default quota of 2 million input tokens per minute, with requests up to 5 million input and 500,000 output tokens per minute possible without extra Anthropic approval, while AWS enforces requests-per-minute limits on its side. AWS's own quota page says defaults can be lower for your account depending on factors such as account age and usage history, that increases are not granted automatically, and that the bedrock-runtime and bedrock-mantle endpoints are tracked against separate quotas, even for the same model.
Foundry: Microsoft's quota page for pay-as-you-go subscriptions, updated 22-09-2026, lists 40 requests per minute, 40,000 input tokens per minute and 8,000 output tokens per minute by default for the current Opus, Sonnet 5 and Haiku 5.5 deployments, and 80 requests, 80,000 input and 16,000 output tokens per minute for Sonnet 4.6, Sonnet 4.5 and Haiku 4.5. Fable 5.1 and Fable 5 show a default of 0. The Enterprise and Free Trial tables did not load for us. Increases go through a request form and are not guaranteed. Foundry also does not send Anthropic's standard rate limit headers, so you monitor with Azure tools.
Vertex AI: the navigation links a "Quotas for Anthropic Claude models" page, but we could not read it, so we state no numbers.
A 40-request-per-minute default means a busy internal tool can hit a 429 within a day, so build exponential backoff in before you launch. For the production side, our LLM operations skills cover monitoring and evaluation.
How do you authenticate on each platform?
Bedrock: for the Messages API endpoint, Anthropic's page lists a Bedrock service role (recommended), IAM assumed roles with a 12-hour maximum session, and bearer tokens with a 12-hour maximum as the least preferred option. The older integration also accepts standard AWS credentials or a bearer token in AWS_BEARER_TOKEN_BEDROCK. Vertex AI: use gcloud application default credentials and pass your project and region to the client, for example region "global". Foundry: an API key in the api-key or x-api-key header, or a Microsoft Entra ID token, which Anthropic's page says typically expires after one hour. Your role needs Foundry User or Cognitive Services User. The Mythos models support Entra ID only.
We would use short-lived, role-based credentials on all three and keep long-lived keys for exploration, which matches what AWS and Anthropic say about Bedrock keys. For the Anthropic API key route see how to get a Claude API key. Our security skills help review IAM policies and secrets handling, and devsecops expert covers CI/CD, infrastructure as code and Kubernetes pipelines.
How do you run Claude Code on Bedrock?
Set CLAUDE_CODE_USE_BEDROCK=1 and an AWS region, or run /setup-bedrock inside Claude Code to use the wizard. The Claude Code Bedrock documentation says Claude Code resolves the region from AWS_REGION, then AWS_DEFAULT_REGION, then your AWS profile, then us-east-1, and that you only need to submit the Anthropic use case form once per AWS account. It uses the Invoke API, not the Converse API, and the WebSearch tool is not available on Bedrock.
Pin your models. Without pinning, the opus alias resolves to Opus 5.5 and the sonnet alias to Sonnet 4.5 on Bedrock, and the page warns that an unpinned primary model bills at the Opus rate. Pin with ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL and ANTHROPIC_DEFAULT_HAIKU_MODEL, and use a geography prefix such as us. or eu. on the inference profile ID. The IAM policy needs bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, bedrock:ListInferenceProfiles and bedrock:GetInferenceProfile. If you use the Mantle endpoint, set CLAUDE_CODE_USE_MANTLE=1 and grant the separate bedrock-mantle actions.
The most common failure is an error saying on-demand throughput isn't supported, which the page fixes by passing an inference profile ID instead of the base model ID. Our DevOps and cloud skills cover the infrastructure side of a rollout, and the env doctor skill finds why a project does not start by checking runtimes, variables, ports and builds. For spend, read how to reduce Claude Code token usage. We did not run Claude Code on Vertex or Foundry for this article, so we state no setup for those.
Which platform should you choose?
Our rule: start from the cloud where your procurement, billing and identity already live, then check the feature gaps above against your design. These are our judgments, not measurements.
If you are on AWS, we would start with the Claude in Amazon Bedrock endpoint, because it uses the same request shape as Anthropic's API, and we would keep the older InvokeModel integration only for Opus 4.6 and earlier. If you are on Google Cloud, Vertex AI is the one where a US and EU multi-region endpoint is documented, which is the point we would show a compliance team first. If you are on Azure, we would deploy the Azure-hosted variant when you need Data Zone processing, and we would steer you away from Foundry for Fable 5.1 or the Mythos models, because the first exists only in the Anthropic-hosted variant and the second is restricted.
What would change our order: a published per-model price table on all three, a documented EU data zone on Foundry, or Message Batches arriving on any of them. If you need the full Anthropic feature set, such as the Files API, Agent Skills or Managed Agents, we would call the Anthropic API directly, because all three clouds list these as missing in at least one deployment type. Whatever you pick, check every figure above against the provider's current page before it goes into a contract or a budget.
Next step: pick one model and one cloud, run a week of your real traffic, and compare the invoice with the Anthropic API price in the table.
Written by Samuel Rose, Founder of Agensi.