Put production model traffic on the cloud where your committed-spend agreement already sits, and default to Amazon Bedrock if you have no commit or have commits on more than one cloud. All three platforms now serve Claude, and list per-token rates match across them, so a price comparison rarely separates them. Two things do. Bedrock is the only one where both Claude and OpenAI's GPT-6 family are generally available with the cloud provider processing the data. Vertex AI is the only place to buy Gemini. Microsoft Foundry is the strongest platform for GPT, with model-independent reserved capacity, and the weakest for Claude, which it sells as a third-party Marketplace product.
The loser for Claude-heavy workloads is Foundry. The loser for Gemini-heavy ones is everything except Vertex. Prices and catalogues below were checked on each vendor's live documentation or price API on October 25, 2026, unless a different date is given.
| Amazon Bedrock | Google Vertex AI (Gemini Enterprise Agent Platform) | Microsoft Foundry | |
|---|---|---|---|
| Claude | Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 and older; AWS runs inference | Same line-up; Google runs inference and bills | Same line-up; Anthropic is the processor; Fable and Mythos only "Hosted on Anthropic" |
| GPT (proprietary) | GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol, GPT-6 Luna, GPT-5.x | GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna in private preview since October 21, 2026 | Full Azure OpenAI line-up, sold by Microsoft |
| Gemini | Gemma open weights only | Only home for Gemini Pro and Flash | No |
| Open weights | Llama, Mistral, DeepSeek, Qwen, Kimi, GLM, gpt-oss, Nemotron | Model Garden | Llama, Mistral, DeepSeek, NVIDIA, Hugging Face |
| Commit drawdown | Bedrock is an AWS service; Marketplace treatment of Claude is set by your EDP | Google says 100% of eligible Marketplace spend may count | Anthropic says Claude is MACC-eligible |
| Reserved capacity for newest Claude | Account-team quote only | Not available (needs a regional endpoint) | Not available (no PTU deployment type) |
| Published uptime SLA | 99.9% per region | 99.5% on Gemini's generateContent methods | 99.9% on Azure OpenAI; partner models vary |
| Verdict | Default pick; widest frontier catalogue | Pick for Gemini; weak for reserved Claude | Pick for GPT; loser for Claude |
What Do Bedrock, Vertex AI and Foundry Actually Serve?
Bedrock has the broadest frontier catalogue of the three, Vertex is the only one with Gemini, and Foundry is the only one where GPT is a first-party Microsoft product. Claude is on all three, but with different operators behind it.
Amazon Bedrock. AWS's model catalogue lists the full Claude 5.x line (Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1, Mythos 5.1 and older), OpenAI's GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol and GPT-6 Luna plus the GPT-5.x models, and open weights from Meta, Mistral, DeepSeek, Qwen, Moonshot (Kimi K3), Z.ai (GLM 5.3) and NVIDIA, plus xAI's Grok 4.7. Google appears only as Gemma 4 open weights. Per Anthropic's Bedrock guide, Claude's global endpoint runs from 27 AWS regions, and geographic inference profiles exist for the US, EU, Japan and Australia. One gap matters for regulated buyers: "For Claude Fable 5.1, regional endpoints are currently available in us-east-1 only."
Google Vertex AI. Google renamed it Gemini Enterprise Agent Platform this year, though most buyers still search for Vertex. It is the only one of the three that sells Gemini. Anthropic's Google Cloud guide lists Claude models up to Fable 5.1, with Google as the operator and the biller. Endpoints are global, multi-region (us or eu) or regional, but "Regional endpoints serve Claude Sonnet 4.6 and earlier models only." On October 21, 2026, Google added GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to Model Garden, in private preview for select enterprise customers.
Microsoft Foundry. Microsoft splits its catalogue in two. Its overview says models "sold by Azure" come with Microsoft support and "enterprise-grade service level agreements," while models from partners, Anthropic named as the example, "are supported by their providers, with varying levels of SLA and support structures" and are billed through Azure Marketplace. Anthropic's Foundry guide adds that Claude deploys only as Global Standard or US Data Zone Standard, that Sonnet 5.5 is Global only, and that Fable 5.1 and the Mythos models are available only "Hosted on Anthropic," on Anthropic's own infrastructure.
For an EU buyer, the Foundry entry is the one that bites: Claude on Foundry has no EU data zone. Claude Fable 5.1 on Bedrock gets regional endpoints only in us-east-1. On Vertex, the eu multi-region endpoint is the only EU option for any Claude model newer than Sonnet 4.6.
Why Token Prices Don't Settle It
List prices for the same model are close enough across the three that they should not drive the decision. Anthropic prices Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens on its own pricing page. On Foundry, Anthropic rates usage "at standard per-model, per-feature rates (same as Claude API pricing)" and converts it to Claude Consumption Units at $0.01 each, so the Foundry price is the API price. For Bedrock and Google Cloud, the same page sends you to each cloud's price list and states that "Regional and multi-region endpoints include a 10 percent premium over global endpoints." Bedrock's pricing page loads its Claude table by script, so we could not capture those rows as text. Third-party trackers such as LLM Reference list Sonnet 5.5 on Vertex at the same $2 input rate. Confirm the per-token line in your own console before you sign anything.
Take one defined workload: Claude Sonnet 5.5, 1 billion input and 100 million output tokens a month, no caching, no batch.
| Path | Monthly list cost | Source of the rate |
|---|---|---|
| Foundry, Global Standard | $3,000 | Anthropic API rate, per Anthropic |
| Foundry, US Data Zone Standard | $3,300 | Equivalent to inference_geo: "us", 1.1x |
| Bedrock or Vertex, global endpoint | Check console; trackers show parity | Each cloud's own price list |
| Bedrock regional or Vertex multi-region | Global rate plus 10% | Anthropic's cloud pricing note |
A $300 difference between global and in-region routing is the largest gap you can document here, and it shows up on every platform. Against a $5 million annual commit, which invoice this workload's $36,000 a year lands on matters far more than any per-token gap.
Does Model Spend Draw Down Your Cloud Commit?
On Foundry, Anthropic states it outright; on Google, Marketplace spend generally counts; on AWS, the answer is in your private EDP terms and nowhere public. This is usually the deciding factor, so get it in writing.
- Azure MACC: Anthropic's Foundry partner page says "Claude models in Microsoft Foundry are eligible for Microsoft Azure Consumption Commitment (MACC)" and that usage "counts toward your MACC." GPT models sold by Azure bill on Azure meters as first-party consumption services, per Microsoft's overview above. Both draw down. The cost of Claude's MACC eligibility is that it arrives as a Marketplace product under Anthropic's terms.
- Google commit: Google's commit drawdown page (updated October 22, 2026) says "100% of spend on Google Cloud Marketplace Services" may count toward minimum commitments, excluding services not running on Google Cloud, Maps, Ads and professional services. Claude on Vertex is invoiced by Google Cloud itself, per Anthropic. Gemini is a Google service and draws down like any other.
- AWS EDP: AWS does not publish how Marketplace purchases count toward an Enterprise Discount Program. Anthropic also sells Claude Platform on AWS, which bills through AWS Marketplace in Claude Consumption Units. Anthropic's pricing page warns that "Discounts cannot be applied retroactively to usage incurred before your private offer is accepted." Ask your AWS account team in writing whether Claude on Bedrock appears as a Bedrock service charge or as a Marketplace charge on your bill. Then ask what share of Marketplace spend your EDP counts.
Steel-man the other side: the hyperscalers treat a large AI commitment as leverage, and a buyer who splits model spend across three clouds gives up that negotiating weight on all three. That is the argument for consolidating. It holds until you need a model one platform does not sell. Our coverage of AWS and Azure DMA duties landing mid-commit covers the renewal timing side of this.
How Does Provisioned Capacity Compare?
Azure's PTUs are the most flexible reserved capacity of the three, and none of the platforms will reserve capacity for the newest Claude at a published price. Our provisioned throughput break-even analysis has the full math; the terms that decide the platform choice are below.
Azure PTUs. The Azure retail price API lists a Global PTU reservation at $260 for one month and $2,652 for one year in East US, and Data Zone or Regional at $286 and $2,916. Microsoft's provisioned throughput guide calls PTUs "model-independent," so a reservation follows you to the next GPT release. It also says "Reservations don't guarantee capacity" and tells you to deploy first and buy second. PTUs cover models sold by Azure. Claude on Foundry has no provisioned deployment type at all.
Vertex Generative AI Scale Units (GSUs). Google's consumption options page offers 1-week, 1-month, 3-month and 1-year terms, with the year "priced at a 26% discount compared to 1-month." On October 22, 2026, Google's pricing page listed global GSUs from $7.14 an hour on a 1-week term down to $2.74 on a 1-year term. For Claude, Anthropic's guide says "provisioned throughput needs a regional endpoint," and regional endpoints stop at Sonnet 4.6. You cannot reserve Opus 5.5 or Sonnet 5.5 capacity on Vertex today.
Bedrock Provisioned Throughput. Model units come on no commitment, 1-month or 6-month terms, and for Claude the pricing page says: "For Provisioned Throughput pricing, please reach out to your account team." AWS's modification guide says "You can't change the model for a Provisioned Throughput associated with a base model." The independent DoiT guide to Bedrock costs calls low-utilization provisioned throughput "the single most expensive mistake on Bedrock."
If reserved capacity on the newest Claude is a hard requirement, Bedrock is the only one of the three where you can even get a quote.
What Do the Uptime SLAs Cover and Exclude?
All three publish availability SLAs, none covers model quality or throttling, and only Bedrock's clearly applies to every model it hosts.
- Bedrock: the Bedrock SLA pays 10% credit below 99.9% monthly uptime per region, 25% below 99.0% and 100% below 95.0%. Exclusions include customer actions such as scaling provisioned throughput, and failures "caused by underlying software that leads to repeated model crashes or an inoperable model." The SLA was last updated October 4, 2023, well before most of today's catalogue existed.
- Vertex: Google's Online Inference SLA sets 99.5% monthly uptime for the
generateContentandstreamGenerateContentmethods, with credits of 10%, 25% and 50% below 99.5%, 99.0% and 95.0%. Those are Gemini's methods. Claude requests on Vertex go to:rawPredictand:streamRawPredict, per Anthropic's guide, so ask Google which SLA covers them before you count on one. - Foundry: Microsoft's SLA covers Azure OpenAI at 99.9% monthly uptime, and provisioned deployments add a latency commitment on token generation, according to Redress Compliance's reading of the SLA. Microsoft's own overview says partner models carry "varying levels of SLA." For Claude, that means Anthropic's terms, not Microsoft's.
Who Processes Your Prompts on Each Platform?
The platform you choose decides who the data processor is, and for Claude the three answers are different.
- Bedrock: Anthropic's guide says Claude runs "on AWS-managed infrastructure with zero operator access (Anthropic personnel have no access to the inference infrastructure)." AWS's abuse detection page says Bedrock does not store inputs or outputs by default, with named exceptions: classifier-flagged traffic on the GPT models it hosts, and all traffic on Claude Fable 5 and 5.1, kept for up to 30 days.
- Vertex: for Google-managed Claude, "Google Cloud governs data handling," per Anthropic, and Anthropic's retention page says that on Bedrock and Google Cloud "the cloud provider is the data processor."
- Foundry: for GPT, Microsoft's data privacy page says models sold by Azure do not interact with services run by their providers, OpenAI included. For Claude, Anthropic's Foundry guide says "Anthropic acts as an independent processor for Microsoft," and that on Azure-hosted deployments "only usage metadata and content flagged by Anthropic's safety systems egress to Anthropic." For Fable 5.1, which is only Hosted on Anthropic, inference runs on Anthropic's infrastructure.
If your data protection team approved "Azure" as the processor, Claude on Foundry needs a second review. Our LLM data residency guide covers the enforcement controls on each cloud.
Who Should Not Pick Each Platform
Amazon Bedrock: skip it if Gemini is your primary model, since Bedrock offers only Gemma. Skip it if you need Fable 5.1 pinned to a single EU region. And skip it if your EDP excludes or caps Marketplace spend and your account team cannot confirm in writing how Claude is billed.
Google Vertex AI: skip it for production GPT workloads while the OpenAI models are in private preview. Skip it if you need reserved capacity on any Claude model newer than Sonnet 4.6. And skip it if you need a published uptime commitment that names the Claude endpoints.
Microsoft Foundry: skip it as your Claude platform if you need EU processing, reserved Claude capacity, Microsoft as the processor, or Fable 5.1 on Azure infrastructure. Claude on Foundry also lacks the Message Batches API, and its Azure-hosted deployments return a 400 error for code execution, Agent Skills and the Files API, per Anthropic's guide.
Anyone with a heavy multi-model mix: do not consolidate on one platform if you run Gemini and GPT in production at volume. No single platform sells both at general availability. Put each model on its home cloud and route between them with a gateway such as LiteLLM or Kong AI Gateway.
The Criteria That Predict Regret
Five questions decide this better than any feature table.
- Where is your largest unspent commit, and how many months are left on it? Model spend that lands there is close to free money. Spend that misses it can mean a true-up at renewal.
- Which model family carries 80% of your tokens? Gemini forces Vertex. GPT at scale favours Foundry for PTU portability. Claude works anywhere, but Bedrock gives you AWS as processor and a path to reserved capacity.
- Does your rule say "processed in region," and which region? Foundry's Claude has no EU zone, and Vertex's newest Claude has no single-region endpoint.
- Will you need reserved capacity within 12 months? Only Azure both publishes reservation prices and lets the reservation follow you to a new model. Only Bedrock will quote reserved capacity on the newest Claude.
- Who approved the processor? Claude on Foundry puts Anthropic in that seat.
What changes the answer: a Google commit with more than a year left and a Gemini-first roadmap flips the default to Vertex. A MACC that will under-run, with GPT as your main model, flips it to Foundry.
What to Do Before You Commit Traffic
This Week:
- Pull each cloud's commit balance and months remaining from finance. Rank the clouds by unspent commit.
- Split last quarter's model tokens by family (Claude, GPT, Gemini, open weights).
- Ask your AWS account team in writing how Claude on Bedrock is billed against your EDP.
This Month:
- Confirm the per-token rate for your main model in each cloud's console, at both global and in-region routing, and write down the date.
- Send the SLA question to Google (which SLA covers Claude's
rawPredictcalls) and to Microsoft (which SLA covers Claude on Foundry). - Have your data protection lead sign off on the processor for each path, by name.
Before Renewal:
- Put the drawdown treatment of each model family into the order form as an exhibit.
- If you need reserved Claude capacity, get Bedrock's tokens-per-minute per model unit and its price in writing before you size anything.
The Bottom Line
Model platforms have converged on catalogue and on price: Claude and the major open-weight families run on all three, and each charges the same 10% for in-region Claude routing. They still differ in the contract: who processes the prompt, what the SLA names, whether reserved capacity follows the model, and whether the spend counts against a commitment you already signed. Our model deprecation comparison adds one more contract term to check, since each cloud sets its own retirement dates.
Start with the finance report that shows your unspent commit by cloud.
Continue Reading
- Annual Azure and Google AI reservations break even at 84% and 75%
- LLM Data Residency: Which Providers Actually Keep Data In Region
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic
- AWS and Azure DMA Duties Land Midway Through a 3-Year Commit
- Model Router Buyer's Guide: Buy Failover, Not Judgment
- Fable 5 Needs Retention On. ZDR Was Never Zero.
- Claude vs GPT vs Gemini: Stop Comparing Per-Token Prices
- AI Gateway vs API Management: No Managed Gateway Caps Your Spend