Amazon Bedrock: which models stay in the EU, at what price

TL;DR

Amazon Bedrock is AWS’s managed API for foundation models from 18 providers. For EU data residency, only Claude, Amazon Nova and a handful of older, smaller or specialised models offer an EU geo profile. The newest OpenAI, xAI and Moonshot models run in Europe only via global routing. Several open-weight models run in a single EU region instead.

What is Amazon Bedrock and how does it work?

Amazon Bedrock is a fully managed AWS service that gives you one API to foundation models from Anthropic, Amazon, OpenAI, Mistral, Meta and 13 other providers, billed on your AWS invoice. You never manage GPUs. You pick a model ID or an inference profile, call it from an AWS region, and pay per token or per reserved capacity. Several apps can share one key and one EU route through a LiteLLM gateway.

AWS itself describes Bedrock as a service that “provides secure, enterprise-grade access to high-performing foundation models” and counts “100+ foundation models” in the Bedrock overview. The model catalogue lists 18 providers as of 19 September 2026.

Three concepts decide how every request behaves.

1. Endpoint. Bedrock has two inference endpoints, documented on the endpoints page:

	bedrock-runtime	bedrock-mantle
APIs	InvokeModel, Converse, Chat Completions, Responses, Messages	Responses, Chat Completions, Messages
Cross-region inference	Yes (geo and global profiles)	No, always single region
Only here	Guardrails, prompt routing	Server-side tools, background jobs, Projects, Workspaces
IAM action	bedrock:InvokeModel	bedrock-mantle:CreateInference
Price per token	Identical	Identical

AWS recommends bedrock-runtime for new applications. Both run on the same Mantle inference engine, so the choice is about features, not quality.

2. Inference profile. This is where residency is decided. In-Region keeps a request in the region you call. A Geo profile (prefix eu., us., jp., au.) routes across a fixed set of regions in one geography, and AWS states that a geo profile’s destination list “will never change”. A Global profile (prefix global.) can route to any commercial AWS region and grows as AWS adds regions.

3. Consumption mode. You pay in one of five ways:

• Standard on-demand: pay per token, no commitment. The default.
• Priority and Flex tiers: set service_tier per request. The pricing page lists Priority at a 75% premium and Flex at a 50% discount, where a model supports them. Claude Sonnet 5 supports Standard only.
• Batch: upload JSONL to S3, get results in S3, at a 50% lower price for supported models. No tool calling.
• Reserved tier: 1 or 3 month capacity reservations, at least 100,000 input and 10,000 output tokens per minute, via your AWS account team. The service tier docs state a 99.5% uptime target.
• Provisioned Throughput: hourly billing per model unit with no commitment, 1 month or 6 months. You need it to serve fine-tuned custom models.

Which models on Amazon Bedrock stay in the EU?

Only two families give you broad EU coverage on current models: Anthropic Claude and Amazon Nova. Both offer an eu. geo profile that keeps processing inside EU regions. Most open-weight models instead run in-region in Frankfurt, Stockholm, Ireland or Milan. The newest frontier models from OpenAI, xAI and Moonshot reach Europe only through global routing.

We checked about 40 model cards on 19 September 2026. The table lists one representative model per provider, with the EU price read from the Amazon Bedrock pricing page. “Not verified” means we could not read an EU price for that exact model. Prices are USD per 1M input / output tokens, Standard tier.

Provider	Example model	EU geo profile	In-region in the EU	EU price in / out
Anthropic	Claude Sonnet 5	Yes, eu.anthropic.claude-sonnet-5	Only via bedrock-mantle: Ireland, Stockholm	2.20 / 11.00
Anthropic	Claude Opus 5	Yes, eu.anthropic.claude-opus-5	Only via bedrock-mantle: Ireland, Stockholm	5.50 / 27.50
Anthropic	Claude Fable 5.1	No, global only	No	10.00 / 50.00 (global)
Amazon	Nova 2 Lite	Yes, eu.amazon.nova-2-lite-v1:0	No	0.429 / 3.597 (from Frankfurt)
Amazon	Nova Pro	Yes, eu.amazon.nova-pro-v1:0	London only (not EU)	1.05 / 4.20 (from Frankfurt)
Mistral AI	Pixtral Large	Yes, eu.mistral.pixtral-large-2502-v1:0	No	2.00 / 6.00
Mistral AI	Devstral 2 123B	No	Frankfurt, Stockholm, Ireland, Milan	0.48 / 2.40
Mistral AI	Mistral Large 3	No	No EU region	not offered in the EU
Meta	Llama 3.2 3B	Yes, eu.meta.llama3-2-3b-instruct-v1:0	No	not verified
Meta	Llama 4 Maverick	No, US geo only	No	not offered in the EU
OpenAI	gpt-oss-120b	No	Frankfurt, Stockholm, Ireland, Milan	0.15 / 0.60 (Stockholm), 0.20 / 0.79 (Frankfurt)
OpenAI	GPT-5.6 Sol	No, US geo and global	No	4.00 / 20.00 (global)
OpenAI	GPT-6 Astra	No, global only in the EU	No	10.00 / 50.00 (global)
xAI	Grok 4.6	No, global only in the EU	No	2.00 / 6.00 (global)
Moonshot AI	Kimi K3	No, global only in the EU	No	3.00 / 15.00 (global)
DeepSeek	DeepSeek V3.2	No	Stockholm	0.74 / 2.22
Qwen	Qwen3 235B A22B	No	Frankfurt, Stockholm, Ireland, Milan	not verified
Z.AI	GLM 5	No	Stockholm	1.20 / 3.84
Google	Gemma 4 31B	No	Frankfurt (bedrock-mantle only)	0.17 / 0.48

The residency and EOL split for those Meta rows is in Llama vs Claude for European companies.

Three findings matter for European buyers:

1. No EU profile for the newest non-Claude frontier models. GPT-6 Astra, GPT-5.6 Sol, Grok 4.6 and Kimi K3 all route globally from every EU region. Our Kimi K3 and GPT-6 Astra analysis covers what that means in detail.
2. The source region can change the price. Claude costs the same from Frankfurt, Ireland and Stockholm. Nova and gpt-oss do not: gpt-oss-120b is 25% cheaper on input in Stockholm than in Frankfurt, and Nova 2 Lite’s EU profile is about 13% cheaper when called from Ireland.
3. “EU profile” depends on where you call it. Nova Pro’s EU profile called from Tel Aviv adds Tel Aviv as a destination, and Claude Haiku 4.5’s profile called from London or Zurich adds those non-EU regions. The same rule applies to Claude Haiku 5.5 on Bedrock: from Frankfurt the destinations stay Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. Always call from a region in an EU member state.

For the Claude-specific setup, including IAM policies that block global routing, read our Claude on AWS Bedrock EU guide. Token rates for DeepSeek V3.2 in Stockholm, and the comparison with Claude and self-hosting, are in DeepSeek pricing 2026.

What does Amazon Bedrock cost? Bedrock pricing in two examples

On-demand Bedrock pricing is per token, with separate input and output rates per model, profile type and sometimes source region. EU geo and in-region rates for Claude are 10% above global. Batch halves the price for supported models. There is no platform fee on top: you pay the model rate and normal AWS charges for S3, logs and networking.

Claude models are sold through AWS Marketplace, so the Sonnet 5 model card notes that charges appear under the model provider in Cost Explorer, not under Amazon Bedrock. Budget alerts must cover both. Whether that Marketplace spend retires your AWS commitment, and how private offers work, is covered in our guide to buying Claude through AWS Marketplace.

Example 1: customer-support assistant on Claude Sonnet 5, EU profile. 20M input and 4M output tokens per month.

Line	Calculation	USD per month
Input	20 × 2.20	44.00
Output	4 × 11.00	44.00
Total, no caching		88.00
Same, with 15M input tokens served from cache (0.22)	5 × 2.20 + 15 × 0.22 + 44.00	58.30 plus cache writes

Prompt caching matters more than the 10% residency premium. The premium on this workload is 8 USD a month. Breakpoints, TTLs and the EU setup sit in Claude prompt caching in the EU.

Example 2: classifying 100M tokens of documents per month, 5M output tokens. Here the model choice changes cost by a factor of more than seven.

Model and location	Calculation	USD per month
gpt-oss-120b, in-region Stockholm	100 × 0.15 + 5 × 0.60	18.00
gpt-oss-120b, in-region Frankfurt	100 × 0.20 + 5 × 0.79	23.95
Nova 2 Lite, EU profile from Ireland	100 × 0.374 + 5 × 3.157	53.19
Claude Haiku 4.5, EU profile	100 × 1.10 + 5 × 5.50	137.50

Our recommendation: route bulk classification to an in-region open-weight model and keep Claude for the steps where quality decides the outcome. For a cross-cloud price view of Claude alone, see our Claude API pricing comparison for the EU. If your staff need seats in the Claude apps rather than tokens, our Claude pricing plans overview lists every plan.

How does Amazon Bedrock protect your data?

On Bedrock, AWS is your processor, not the model vendor. AWS documents that model providers have no access to Bedrock deployment accounts, logs, prompts or completions. You control retention per region, including zero data retention for models that allow it. The main exceptions are Claude Fable models, which require up to 30 days of AWS-internal retention. Mode none in Frankfurt, the SCP and the features that still store data are in Claude Zero Data Retention in the EU.

The data protection page describes model deployment accounts operated by AWS per region and provider. TLS 1.2 is required and 1.3 recommended. The data retention page defines the modes:

Mode	Effect
none	Zero data retention: nothing written to durable storage, nothing shared with the provider
default	Model default applies, which can include abuse-detection storage
aws_review	Up to 30 days of retention inside AWS for review; required for Claude Fable 5 and 5.1

Four operational points decide whether this holds up in a GDPR review:

• Set retention per region. The setting does not propagate. Enforce none organisation-wide with the bedrock:DataRetentionMode condition key in an SCP.
• Cross-region storage follows the destination. With a geo or global profile, retained data sits in the region that processed the request. With the EU profile, that is still an EU region.
• Opt-in regions. AWS warns that cross-region profiles can route to opt-in regions you never enabled, and that prompts may be stored there for abuse detection. Check the destination list before go-live.
• Audit trail. CloudTrail logs each request in the source region and records the actual processing region in additionalEventData.inferenceRegion.

If the VPC must have no public egress, create an interface endpoint in Frankfurt; service names, the endpoint policy and why the EU profile still matters are in Bedrock PrivateLink for Claude. If you run coding agents, the same rules apply to Claude Code; our Claude Code with Bedrock setup shows the environment variables that pin it to the EU profile. If many developers share one Bedrock account, our AI gateway vs LLM gateway comparison shows how a gateway holds the AWS credential and maps every model to the eu. profile.

When is Amazon Bedrock the right choice?

Choose Amazon Bedrock when your workloads already run on AWS, when you want several model families behind one IAM and billing model, or when you need open-weight models in a single EU region. Choose Google Cloud when you need Claude Fable 5.1 inside the EU; the residency, 30-day retention and prices are in our Claude Fable 5.1 Europe guide. Self-hosting only pays off at high, steady volume or with strict model control needs. If nobody in house will own IAM, logging and model lifecycle, our guide to managed AI services for Claude shows what an external operator takes over on Bedrock and what stays with you.

Criterion	Amazon Bedrock	Google Cloud (Claude)	Self-hosting
EU residency for Claude Opus 5 / Sonnet 5	EU geo profile	EU multi-region endpoint	n/a (Claude weights are not public)
Claude Fable 5.1 in the EU	No, global only	Yes, EU multi-region at 11 / 55 USD	n/a
Claude Sonnet 5 EU price	2.20 / 11.00	2.20 / 11.00	n/a
Open-weight models in one EU region	Yes (gpt-oss, Devstral, DeepSeek, Qwen, GLM, Gemma)	Not covered in this guide	Yes, any model
Guardrails, Knowledge Bases, Agents	Built in	Own equivalents	Build yourself
Operations effort	Low	Low	High (GPUs, scaling, patching)

How to run a Managed Knowledge Base in Frankfurt or Ireland with AgenticRetrieveStream is in agentic RAG with Claude in the EU.

Google’s EU multi-region strictly covers EU member states, so it excludes the UK and Switzerland. The Google platform itself, the Agent Platform rename and the EU multi-region prices are in Vertex AI explained. Our Claude on Vertex AI in Europe guide explains the setup, and the GDPR comparison of all Claude providers adds Microsoft Foundry and the Anthropic API. What Azure AI Foundry is, which Claude models it hosts and why it has no EU data zone is in Azure AI Foundry explained. Classification traffic that should not burn Claude output tokens is a different job; our Jev TypeSafe decision model guide covers that US-hosted path and when a self-hosted classifier is safer.

Bedrock also has a middle ground between managed models and your own GPUs: Custom Model Import. In Frankfurt it costs 0.07144 USD per Custom Model Unit per minute, billed in 5-minute windows. A Llama 3.1 70B model needs 8 units, which is about 34 USD per active hour, or roughly 25,000 USD a month if it never scales to zero. That is the benchmark any self-hosting business case has to beat, for example the one in our DeepSeek V4.1 Flash self-hosted guide. For market numbers across providers, see our statistics page.

How do you make your first Amazon Bedrock call from an EU region?

You need an AWS account, IAM permissions for Bedrock in an EU region, and either AWS credentials or a Bedrock API key. Then you call the Converse API with an EU inference profile from eu-central-1. The whole path takes about 15 minutes if your organisation’s SCPs already allow the EU destination regions.

1. Pick a source region in an EU member state, for example eu-central-1 (Frankfurt) or eu-west-1 (Ireland).
2. Check SCPs. Allow bedrock:InvokeModel* in every destination region of the profile. AWS states that one blocked destination makes requests fail.
3. Grant IAM for the inference profile ARN and the underlying foundation model, and deny aws:RequestedRegion = unspecified to block global routing.
4. Authenticate. Use your normal AWS credential chain, or create a long-term Bedrock API key in the console and export it as AWS_BEARER_TOKEN_BEDROCK. Short-term keys (12 hours), long-term keys (1 to 365 days) and why the EU profile still matters are in Bedrock API keys for Claude in the EU.
5. Verify destinations, then send a test request with the commands below.
6. Set data retention to none in each EU region you use, if your model allows it. Prompts and answers reach your account only after you enable Bedrock model invocation logging in the EU: off by default, bedrock-runtime only, retention is yours to set.

Both commands are single lines and copy-paste ready:

aws bedrock get-inference-profile --region eu-central-1 --inference-profile-identifier eu.anthropic.claude-sonnet-5 --query "models[].modelArn" --output table

aws bedrock-runtime converse --region eu-central-1 --model-id eu.anthropic.claude-sonnet-5 --messages '[{"role":"user","content":[{"text":"Explain Amazon Bedrock in two sentences."}]}]' --inference-config '{"maxTokens":512}'

The first command prints one model ARN per destination region. If a region outside the EU appears, stop.

The same call in Python with boto3 and the Converse API. Converse uses one request shape for all providers, so switching to eu.amazon.nova-2-lite-v1:0 only changes the model ID:

# pip install boto3
import boto3

client = boto3.client("bedrock-runtime", region_name="eu-central-1")

response = client.converse(
    modelId="eu.anthropic.claude-sonnet-5",
    messages=[{"role": "user", "content": [{"text": "Explain Amazon Bedrock in two sentences."}]}],
    inferenceConfig={"maxTokens": 512},
)

for block in response["output"]["message"]["content"]:
    if "text" in block:
        print(block["text"])
print(response["usage"])

Sonnet 5 thinks adaptively by default, and thinking tokens count toward maxTokens. Raise the limit if replies come back cut off.

Whether Mistral Large 3, Devstral 2 or Claude Sonnet 5.5 stay in the EU is in Mistral vs Claude for European companies. Qwen3 on Bedrock stays in a single Region with no geo profile; price and residency are in Qwen vs Claude for European companies. Gemma 4 has a documented on-demand rate in Frankfurt and no geo profile; that comparison is in Gemma vs Claude for European companies. How to pin Claude data residency on Bedrock, Google Cloud and Foundry is in Claude data residency in the EU. When a private LLM in the EU (your own weights, import into Frankfurt or Gemma 4 in eu-central-1) beats Claude as the default depends on price, licence and residency. When procurement writes sovereignty, not only residency, the five-check list is in sovereign AI in Europe.

FAQ

The answers below stand alone and reflect AWS documentation read on 19 September 2026. Bedrock adds models often, so check the model card before you sign a data processing agreement.

What is Amazon Bedrock in one sentence?

Amazon Bedrock is a fully managed AWS service that gives you API access to foundation models from 18 providers, including Anthropic Claude, Amazon Nova, OpenAI and Mistral, billed through your AWS account without running any GPU infrastructure yourself.

How much does Amazon Bedrock cost?

There is no base fee; you pay per token. In the EU, Claude Sonnet 5 costs 2.20 USD per million input tokens and 11 USD per million output tokens via the EU profile. Open-weight models are far cheaper: gpt-oss-120b costs 0.15 / 0.60 USD in Stockholm. Batch cuts supported models by 50%.

Amazon Bedrock vs Google Cloud: which is better for Claude in the EU?

For Opus 5 and Sonnet 5 they are equal on price and both keep data in the EU. Bedrock wins if you want other model families and AWS-native guardrails. Google wins if you need Claude Fable 5.1 with EU residency, because Bedrock only offers it through global routing.

Does Amazon Bedrock keep my data in the EU?

Only if you choose correctly. Use an eu. geo profile or an in-region model called from an EU member state region, and block global profiles in IAM. Models with only a global profile, such as GPT-6 Astra or Kimi K3, can process data anywhere in the world.

Do OpenAI or Anthropic see my prompts on Bedrock?

No. AWS states that model providers have no access to Bedrock deployment accounts, logs, prompts or completions. AWS may retain data for abuse detection or, for Claude Fable models, for up to 30 days of review, but it is not shared with the provider.

Is AWS Bedrock the same as Amazon Bedrock?

Yes. “AWS Bedrock” is the informal name. The official product name is Amazon Bedrock, and both terms refer to the same service, APIs and pricing.

Sources

1. AWS Bedrock docs: What is Amazon Bedrock (19 September 2026)
2. AWS Bedrock docs: Endpoints supported by Amazon Bedrock (19 September 2026)
3. AWS Bedrock docs: Models at a glance (19 September 2026)
4. AWS Bedrock docs: Claude Sonnet 5 model card (19 September 2026)
5. AWS Bedrock docs: Nova 2 Lite model card (19 September 2026)
6. AWS Bedrock docs: gpt-oss-120b model card (19 September 2026)
7. AWS Bedrock docs: Supported Regions and models for inference profiles (19 September 2026)
8. AWS Bedrock docs: Service tiers (19 September 2026)
9. AWS Bedrock docs: Batch inference (19 September 2026)
10. AWS Bedrock docs: Provisioned Throughput (19 September 2026)
11. AWS Bedrock docs: Data retention (18 September 2026)
12. AWS Bedrock docs: Data protection (18 September 2026)
13. Amazon Bedrock pricing (19 September 2026)
14. Google Cloud generative AI pricing (18 September 2026)