Take Vertex AI unless you need to pin a single country — then take Bedrock. Both charge the same 10% residency premium, so price is not the tiebreaker. Vertex keeps structured outputs, the web search tool and the browser use tool that Bedrock drops; it serves the entire current lineup including Fable 5.1, which Bedrock cannot serve from Europe at all; and Google states in writing that its eu multi-region excludes the UK and Switzerland. Bedrock’s advantages are narrower and real: a published TPM quota, provisioned capacity, inference inside the AWS security boundary, and the ability to pin Stockholm or Ireland as a single region — which Vertex cannot do for any current model.
Why there are only two options
Anthropic’s first-party API has no EU endpoint. Its residency product is US-only, priced on claude.com/pricing as “US-only inference… at 1.1x pricing for input and output tokens,” and the geo parameter takes us or global. There is no eu. Claude in Microsoft Foundry has no EU Data Zone either, which this desk documented when mapping every EU-resident API.
So a European team that chose Claude on capability — and many did, because the alternative with EU jurisdiction is Mistral, which is not an agentic-coding substitute — has two procurement paths and no third. This page is the comparison between them.
The feature delta, which has nothing to do with Europe
This is the decision most buyers get wrong, because they evaluate residency and assume the model is otherwise the same Claude. It is not. Both marketplaces ship a reduced feature set, and Bedrock’s is meaningfully smaller than Vertex’s.
| Capability | Claude API | Vertex AI | Amazon Bedrock |
|---|---|---|---|
| Messages API, prompt caching, thinking, tool use, citations | Yes | Yes | Yes |
| Structured outputs | Yes | Yes | No |
| Web search tool | Yes | Yes | No |
| Browser use tool | Yes | Yes | No |
computer_toolset_20260801 / browser_toolset_20260801 | Yes | Yes | No (beta computer-use versions only) |
| Files API, URL inputs | Yes | No | No |
| Code execution, web fetch, advisor tools | Yes | No | No |
| Agent Skills, MCP connector, programmatic tool calling | Yes | No | No |
| Message Batches, Models / Admin / Usage endpoints | Yes | No | No |
| Claude Managed Agents | Yes | No | No |
Server-side fallback (fallbacks) | Yes | No | No |
| 1M-token context on current models | Yes | Yes | Yes |
Verified 10 October 2026 against Anthropic’s Claude in Amazon Bedrock and Claude on Google Cloud pages.
Four rows separate the two, and structured outputs is the one that breaks ports. A service that relies on guaranteed-schema JSON responses runs on the Claude API and on Vertex, and has to be rewritten with prompt-level schema coaxing and validation retries on Bedrock. Budget that as engineering work, not as a config change. The web search and browser use tools are the other two: if your agent searches, Bedrock means wiring your own search backend.
The shared losses matter too. Message Batches is unavailable on both, which removes the 50% batch discount from the marketplace routes entirely — a bigger number than the residency premium on any batchable workload.
What each platform means by “EU”
The two vendors define the boundary differently, and only one of them defines it in writing.
Google states it. Vertex’s data-residency documentation says the EU multi-region “strictly covers data residency within EU member states. Geographies outside the European Union political boundary, including the United Kingdom and Switzerland, are excluded.” Claude is listed as supported on both the US and EU multi-regions.
AWS names a prefix. Anthropic’s Bedrock region table marks eight European AWS regions as carrying the EU inference profile:
| AWS region | Location | In the EU/EEA? | Single-region pinning |
|---|---|---|---|
eu-central-1 | Frankfurt | Yes | No — EU profile only |
eu-central-2 | Zurich | No — Switzerland | No — EU profile only |
eu-north-1 | Stockholm | Yes | Yes (in-region) |
eu-south-1 | Milan | Yes | No — EU profile only |
eu-south-2 | Spain | Yes | No — EU profile only |
eu-west-1 | Ireland | Yes | Yes (in-region) |
eu-west-2 | London | No — United Kingdom | No — EU profile only |
eu-west-3 | Paris | Yes | No — EU profile only |
Both Switzerland and the UK hold adequacy decisions, so a transfer to either is lawful. But it is a transfer to a third country, and a control written as “EU member states” is not satisfied by a response generated in Zurich. AWS does not publish the EU profile’s destination list on that page — it tells you to go and get it: “If you have data residency requirements, call GetInferenceProfile from every source Region that you plan to use and verify all destination Regions in the response… Don’t rely on the geographic prefix in an inference profile ID alone to determine where requests can be routed.”
Two further AWS notes belong in the risk register. Destination regions “can include opt-in Regions”, and “your inference request can be routed to any of the destination Regions in the profile, even if you did not opt-in to such Regions in your account. Your input prompts and output results may be stored in the opt-in Regions for abuse detection purposes.” On the reassuring side, AWS commits that for a geography-tied profile “its destination Region list will never change” — so the footprint is stable once you have measured it. Measure it.
The pinning asymmetry
This is the one axis where Bedrock wins outright, and it is the reason to choose it.
Bedrock can pin a single country — but only two of them. Anthropic’s table marks eu-north-1 (Stockholm) and eu-west-1 (Ireland) as supporting “direct single-region routing without an inference profile.” Frankfurt, Paris, Milan, Madrid, Zurich and London are reachable as EU-profile destinations, not as pinned single regions. If your requirement is literally “all inference in Germany,” Bedrock does not sell that for Claude either.
Vertex cannot pin at all for current models. Anthropic’s documentation is explicit that “specific regional endpoints support Claude Sonnet 4.6 and earlier; newer models use the global or multi-region endpoints.” For Opus 4.7 and everything after it, your EU choice on Vertex is the eu multi-region — Belgium, Netherlands, Germany and France in the residency tables — or nothing. The same constraint removes provisioned throughput, which Vertex’s docs tie to regional endpoints.
So: single-region mandate → Bedrock, Stockholm or Ireland. Geography-level mandate → either, with Vertex’s boundary the better-documented one.
The price: a tie, and a tokeniser that costs more than the tie
Anthropic applies one rule to both platforms: “Regional and multi-region endpoints include a 10% premium over global endpoints,” for Claude Sonnet 4.5 and all later models. There is no price reason to prefer either marketplace.
A reference month of 200 million characters in and 40 million characters out — a mid-size production assistant — converted at Anthropic’s current tokeniser of 2.692 characters per token (74.29M input tokens, 14.86M output tokens):
| Model | Rate (in / out per MTok) | Global endpoint | EU-resident (+10%) | Residency cost/yr |
|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 / $0.50 | $14.86 | $16.34 | $18 |
| Claude Sonnet 5.5 | $2 / $10 | $297.18 | $326.89 | $357 |
| Claude Opus 5.5 | $4 / $20 | $594.35 | $653.79 | $713 |
| Claude Opus 5 | $5 / $25 | $742.94 | $817.24 | $892 |
| Claude Fable 5.1 | $10 / $50 | $1,485.88 | $1,634.47 (Vertex only) | $1,783 |
Prices verified 10 October 2026 against Anthropic’s pricing page, and mirrored in our pricing tracker. A caveat worth knowing before you quote a number to finance: on 10 October neither marketplace’s own price list returned current Claude rows to an automated read — AWS’s Bedrock pricing page rendered Claude 3.5-era rows and Google’s Agent Platform pricing page rendered none. Anthropic’s own page is the authoritative rate card for both marketplaces; price your workload in each vendor’s calculator before committing.
Two arithmetic notes. Haiku 5.5 multiplies every rate by 5x above a 100,000-token prompt, and that threshold is counted in the newer tokeniser, so it arrives sooner than it looks. And the tokeniser itself is the larger line item: at identical $2/$10 per-token rates, EU-resident Sonnet 5.5 costs $326.89 where GPT-6.1 Sol costs $220.00 on the same text — a 49% gap created by tokenisation, not by residency. Our API cost calculator applies the per-vendor ratios.
The retirement schedules are not Anthropic’s
A detail that favours the marketplaces and almost nobody prices in: partner platforms set their own lifecycle dates. Anthropic’s note is that its published dates “apply to the Claude API, Claude Platform on AWS and Microsoft Foundry. Amazon Bedrock and Google Cloud set their own schedules,” and the Vertex documentation repeats it: “lifecycle dates on partner-operated platforms are set by the partner and can differ from the Claude API schedule.”
In practice this means the marketplaces run the longest tails. Claude Opus 4.1 and Sonnet 4 are retired on the first-party API but still served on both Bedrock and Google Cloud; Claude Opus 4 and Haiku 3.5 are retired everywhere except Google Cloud. If you are pinned to an older snapshot and need runway, Vertex has the longest. Dates for every model are in our model retirement tracker — but check the partner’s page, not Anthropic’s, for a marketplace deployment.
Decision table
| Your binding constraint | Buy |
|---|---|
| EU residency, default case | Vertex eu multi-region — more features, documented boundary |
| Structured outputs, web search or browser use in the product | Vertex — Bedrock supports none of the three |
| Claude Fable 5.1 under EU residency | Vertex — Bedrock regional is us-east-1 only |
| Inference pinned to one country | Bedrock, eu-north-1 (Stockholm) or eu-west-1 (Ireland) |
| A published throughput quota or reserved capacity | Bedrock — 2M input TPM default, 5M on request |
| Inference inside the AWS security boundary, AWS billing | Bedrock — zero Anthropic operator access, CloudTrail logging |
| A control that says “EU member states”, audited | Vertex — Google excludes UK and Switzerland in writing |
| Batch processing at the 50% discount | Neither — Message Batches is unavailable on both |
| EU jurisdiction rather than EU location | Neither — see Mistral or Cohere sovereign |
If the real question behind the shortlist is which assistant rather than which endpoint, Claude vs ChatGPT is the capability axis and Claude Code is the agent one; this page is only procurement.
Before you sign
- Query the profile, do not read the prefix. On Bedrock, run
GetInferenceProfilefrom every source region you will call and record the destination list. AWS tells you to; almost nobody does. - Decide whether “EU” means the EEA or the member states. Zurich and London are the difference, and it is a one-line change to your DPIA either way — but only if you noticed.
- Test structured outputs before you pick Bedrock. If the service depends on schema-guaranteed JSON, this is a rewrite, not a migration.
- Confirm provisioned capacity in writing on Vertex. The documentation implies you cannot have it under EU residency for a current model.
- Price it off Anthropic’s page, then re-price it in the marketplace calculator. Neither marketplace’s public rate card was readable on 10 October 2026.
What changed
- 10 Oct 2026: Page created. Every cell verified against Anthropic’s Bedrock and Google Cloud documentation, AWS’s inference-profile guidance and Google’s Vertex data-residency tables on the same day.
Frequently asked questions
Can I use Anthropic's own API with EU data residency?
No. The Claude API is global by default and its first-party residency option is US-only — claude.com/pricing states that 'US-only inference is available at 1.1x pricing for input and output tokens', and the geo parameter accepts us and global with no eu value. Claude under EU data residency is a cloud-marketplace purchase, through Amazon Bedrock or Google Cloud's Agent Platform (Vertex AI), and nothing else. Claude in Microsoft Foundry has no EU Data Zone at all. If your procurement process assumes you will contract with Anthropic directly for an EU-resident endpoint, it is assuming something that is not on sale.
Is Bedrock or Vertex cheaper for EU-resident Claude?
Neither. Anthropic's pricing documentation applies the same rule to both: 'Regional and multi-region endpoints include a 10% premium over global endpoints', covering Bedrock regional endpoints and Google Cloud multi-region endpoints alike, for Claude Sonnet 4.5 and every model since. On a 200M-character-input, 40M-character-output month, EU-resident Sonnet 5.5 costs $326.89 on either platform against $297.18 on a global endpoint. The residency decision is worth about $30 a month on that workload. The platform decision is worth a feature set.
Which Claude models can I actually run in the EU on each platform?
Vertex serves the whole current lineup from its eu multi-region, including Claude Fable 5.1. Bedrock does not: Anthropic's Bedrock documentation states that 'for Claude Fable 5.1, regional endpoints are currently available in us-east-1 only', so Fable 5.1 under EU residency is a Vertex-only purchase today. Claude Mythos Preview is likewise us-east-1 only on Bedrock. Everything else — Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, Haiku 5.5, Haiku 4.5 and the 4.x Opus models — carries the EU inference profile on Bedrock and runs in Vertex's eu multi-region.
Does the Bedrock EU inference profile keep my data inside the EU?
Not necessarily, and AWS says so. Anthropic's region table marks eight European AWS regions as carrying the EU inference profile, two of which are outside the EU and the EEA: eu-central-2 is Zurich (Switzerland) and eu-west-2 is London (United Kingdom). Both hold adequacy decisions, so a transfer is lawful — but it is still a transfer to a third country, and a control that says 'EU member states' is not satisfied by a request served from Zurich. AWS's own guidance is blunt: 'If you have data residency requirements, call GetInferenceProfile from every source Region that you plan to use and verify all destination Regions in the response... Don't rely on the geographic prefix in an inference profile ID alone to determine where requests can be routed.' Google takes the opposite approach and states the boundary in the documentation: its EU multi-region 'strictly covers data residency within EU member states. Geographies outside the European Union political boundary, including the United Kingdom and Switzerland, are excluded.'
What do I lose by running Claude on a marketplace instead of Anthropic's API?
On both platforms you lose the Files API and URL inputs, the code execution and web fetch server-side tools, Agent Skills, the MCP connector, programmatic tool calling, Message Batches, the Models/Admin/Usage endpoints, Claude Managed Agents and server-side fallback. Bedrock additionally drops structured outputs, the web search tool, the browser use tool and the current computer_toolset_20260801 and browser_toolset_20260801 toolsets — all four of which Vertex supports. That gap is the single biggest difference between the two routes and it has nothing to do with Europe: it applies in every region.
Can I buy provisioned capacity for EU-resident Claude?
On Bedrock, yes, and the quota is published: 2 million input tokens per minute by default, raisable to 5 million input and 500,000 output TPM without additional Anthropic approval. On Vertex it is doubtful for current models. Anthropic's documentation says provisioned throughput requires a regional endpoint, that multi-region endpoints 'only support pay-as-you-go traffic', and that specific regional endpoints serve 'Claude Sonnet 4.6 and earlier' — which together mean an EU-resident Opus 5.5 or Sonnet 5.5 on Vertex is a metered purchase, not a reserved one. If reserved capacity under EU residency is a hard requirement, get it confirmed in writing by your Google account team before you design around it.