Anthropic Claude is a family of large language models from the US company Anthropic, which puts safety and transparency at the centre of its model development. The models come in three performance tiers: Opus for complex, long-running work, Sonnet for broad everyday use and Haiku for high volume and low latency. Above them sits the access-restricted Mythos class, with Claude Fable 5.2 as its generally available variant. The current reference point is the Claude 5.5 family with Opus 5.5, Sonnet 5.5 and Haiku 5.5.
For companies in Germany, Austria and Switzerland, the delivery route matters more than the individual model: besides the Anthropic API, Claude runs on AWS Bedrock, Google Cloud’s Gemini Enterprise Agent Platform (formerly Vertex AI) and Microsoft Foundry. EU regions, data residency and retention obligations differ considerably by model and platform. This overview sets out the current Claude models, prices and EU options. As an AI consultancy from Rosenheim we support model selection and GDPR-compliant integration, for example via CompanyGPT and the innFactory AI Gateway.
Key facts at a glance (as of November 2026)
- Anthropic is behind Claude, a public benefit corporation from San Francisco, founded in 2021. The contracting party for customers in the EEA is Anthropic Ireland, Limited in Dublin.
- Claude can be used GDPR-compliantly if the delivery route and contract fit: for business customers, Anthropic acts as a processor with EU standard contractual clauses and does not train on customer data.
- Anthropic does not offer its own EU servers. The Claude API and claude.ai store data in the US. EU processing is available via AWS Bedrock (EU geo profile) and Google Cloud’s Gemini Enterprise Agent Platform (EU multi-region).
- Current recommendation: Claude Opus 5.5 for demanding work, Sonnet 5.5 for everyday use, Haiku 5.5 for high volume; all three without mandatory 30-day retention. For the hardest tasks, Claude Fable 5.2 has been available since 9 November 2026 (USD 8 / 40, with mandatory 30-day retention).
- Procurement in Germany: Through innFactory GmbH as a cloud reseller, Bedrock, Agent Platform and Foundry can be obtained on invoice, combined with the AI Gateway and CompanyGPT.
Current models at a glance
| Model | API ID | Release | Input / output (USD per 1M tokens) | Context | Focus |
|---|---|---|---|---|---|
| Claude Opus 5.5 | claude-opus-5-5 | 22 September 2026 | 4 / 20 | 1M | Complex agentic work, coding, knowledge work |
| Claude Sonnet 5.5 | claude-sonnet-5-5 | 28 September 2026 | 2 / 10 | 1M | Everyday tasks, documents, coding, support |
| Claude Haiku 5.5 | claude-haiku-5-5 | 7 October 2026 | from 0.10 / 0.50 | 1M | High volume, classification, routing, subagents |
| Claude Fable 5.2 | claude-fable-5-2 | 9 November 2026 | 8 / 40 | 1M | Highest capability, with mandatory 30-day retention |
| Claude Mythos 5.2 | claude-mythos-5-2 | 9 November 2026 | 8 / 40 | 1M | Only after vetting (Anthropic’s verification programmes) |
All five models have a June 2026 knowledge cutoff, deliver up to 128K output tokens and use adaptive thinking, whose depth is controlled by the effort parameter. Older models such as Fable 5.1, Opus 5, Sonnet 5, Opus 4.8 and Haiku 4.5 remain available and are still relevant for certain EU setups (see the sections on EU availability and lifecycle).
Claude Opus 5.5
Claude Opus 5.5 is the current Opus model and the first of the Claude 5.5 family. Anthropic positions it for long-running agentic coding and knowledge work and describes it as a model that performs at the level of Fable 5.1 on most work while costing around 40% less to run than Opus 5. Before release it was tested by external evaluators, including METR.
Per Anthropic, Opus 5.5 at maximum effort scores 66.4% on Terminal-Bench 4.0 (Fable 5.1: 55.8%, Opus 5: 52.3%), 54.4% on FrontierCode v1.1, 57.8% on CursorBench 4.0, 1846 Elo on GDPval-AA v2.1, 67.7% on Humanity’s Last Exam with tools and 81.8% on OSWorld 2.0 (partial). Anthropic itself notes that benchmark gaps at this level are a less reliable measure of real-world differences.
Opus 5.5 costs USD 4 input and USD 20 output per 1M tokens, cache reads USD 0.20. It is not a covered model, so zero data retention is possible. Adaptive thinking is always on and the default effort is medium. The fast mode for Opus 5.5 has been generally available on the Claude API and in Microsoft Foundry (hosted on Azure) since 15 October 2026, and via Claude Models on Google Cloud (Anthropic managed) since 22 October; it costs USD 8 / 40. Opus 5.5 ships with safeguards similar to Fable 5.1: cyber and biology classifiers plus a classifier against reasoning extraction; declined requests return stop_reason: "refusal".
Claude Sonnet 5.5
Claude Sonnet 5.5 is the mid-tier model of the Claude 5.5 family and, per Anthropic, a faster, lower-cost complement to Opus 5.5. It is strongest at well-scoped everyday tasks, fixing bugs and creating documents, slides and spreadsheets. For complex, open-ended work requiring sustained judgment, Opus 5.5 remains clearly stronger per Anthropic.
The price matches Sonnet 5 at USD 2 input and USD 10 output per 1M tokens; cache reads cost USD 0.10. Because Sonnet 5.5 needs far fewer tokens for the same work, it costs up to 30% less per task than Sonnet 5 per Anthropic and generates output more than 30% faster. Anthropic cites 70.6% on Terminal-Bench 4.0 (Sonnet 5: 10.3%), 55.5% on CursorBench 4.0 (34.1%), 1844 Elo on GDPval-AA v2.1 (1449) and 80.1% on OSWorld 2.1 (partial, 57.0%), among others.
Sonnet 5.5 is not a covered model and, per Anthropic, is available with zero data retention. The default effort on the Claude Platform is high. It is the first Sonnet model with cyber safeguards like those on Opus 5.5: routine software development is unaffected, while higher-risk cyber tasks visibly fall back to Sonnet 5. Per its system card it is the most robust Sonnet model yet against prompt injection.
Claude Haiku 5.5
Claude Haiku 5.5 is the small model of the Claude 5.5 family, built for high-volume, latency-sensitive work: classification, routing, extraction, summaries and compaction, live customer support, browser automation and subagents supporting Opus 5.5 or Sonnet 5.5. Anthropic describes it as the cheapest, fastest and most capable small model it has released so far.
Pricing is tiered by prompt length: up to 100,000 tokens Haiku 5.5 costs USD 0.10 input and USD 0.50 output per 1M tokens, above that USD 0.50 / 2.50. According to Anthropic, around 90% of previous Haiku requests fall into the lower tier, so on average Haiku 5.5 costs about 75% less than Haiku 4.5. It is the first Haiku model with adaptive thinking and the effort parameter (default medium) and offers a 1M token context instead of 200K.
Per its system card, Haiku 5.5 at maximum effort scores 39.2% on Terminal-Bench 4.0 (Haiku 4.5: 0.0%), 72.4% on the OSWorld 2.1 offline subset (15.7%), 83.7% on SWE-bench Multilingual (67.4%) and 1620 Elo on GDPval-AA v2.1 (735). Haiku 5.5 is not a covered model, so zero data retention is possible. There is no Priority Tier.
Claude Fable 5.2 and Mythos 5.2
Claude Fable 5.2 has been the most capable generally available Claude model since 9 November 2026 and replaces Fable 5.1. Anthropic positions it for especially demanding reasoning and long agentic runs: sessions lasting hours, multi-step research and analyses through to the finished document. Per Anthropic it scores 69.9% on FrontierSWE v2 (Fable 5.1: 56.3%, Opus 5.5: 62.3%), 74.6% on Terminal-Bench-Science 0.1 (52.6%), 51.4% on AutomationBench (31.4%), 74.2% on Humanity’s Last Exam with tools (65.6%) and 41.3% on the Remote Labor Index (17.9%). Per Anthropic, it generates output up to 30% faster than Fable 5.1.
Fable 5.2 is 20% cheaper than its predecessor: USD 8 input and USD 40 output per 1M tokens, cache reads USD 0.20, and the fast mode costs USD 16 / 80. Per Anthropic, the safety classifiers block around 55% fewer requests than on Fable 5.1. High-risk cyber tasks such as penetration testing and exploit development as well as most research biology tasks remain blocked; these requests are answered by Opus 4.8 and Opus 5 respectively. For most workloads Anthropic still recommends Opus 5.5 as the starting point and Fable 5.2 only when Opus 5.5 is not sufficient even at higher effort.
The decisive point for a GDPR assessment: like its predecessors, Fable 5.2 is a covered model. Prompts and outputs are retained for 30 days on every platform, and zero data retention is only available with explicit authorization by Anthropic. Per Anthropic, eligible Enterprise customers can use Fable 5.2 and Mythos 5.2 with ZDR, even before they adopt the Enterprise Frontier Safeguards (EFS). EFS monitors prompts and responses that remain stored in the customer’s infrastructure and, per Anthropic, is currently in private beta. In the EU, Fable 5.2 is available from day one via the Europe multi-region of the Gemini Enterprise Agent Platform; on AWS Bedrock, as with Fable 5.1, only the global. and us. profiles exist.
Claude Mythos 5.2 is the same model with less restrictive cyber and biology safeguards and, per Anthropic, the strongest Claude model for cybersecurity: on Anthropic’s internal CVE Rediscovery eval it finds 14.4% more known vulnerabilities than Mythos 5.1. Access is granted only to organizations that Anthropic has vetted via the Cyber Verification Program or the Life Sciences Verification Program; former participants in Project Glasswing go through the “Specialized Access” tier. Whether EU organizations receive access is not publicly documented. Claude Enterprise customers use Mythos 5.2 indirectly, even without verification, through Claude Security, which scans codebases for vulnerabilities and proposes patches for human review.
Previous models
- Claude Opus 5 (
claude-opus-5, 24 July 2026, USD 5 / 25): The Opus flagship until Opus 5.5. It remains relevant because it runs in-region in Ireland and Stockholm on AWS Bedrock. - Claude Sonnet 5 (
claude-sonnet-5, 30 June 2026, USD 2 / 10): Predecessor of Sonnet 5.5, also with in-region hosting in Ireland and Stockholm via the bedrock-mantle endpoint. Per Anthropic 34.6% on Humanity’s Last Exam without tools. - Claude Opus 4.8 and 4.7 (May and April 2026, USD 5 / 25): Still generally available, superseded by Opus 5.5 for new projects. Per Anthropic, Opus 4.8 scores 69.2% on Terminal-Bench 2.
- Claude Opus 4.6 and Sonnet 4.6: The newest Opus and Sonnet models with a regional endpoint in Belgium (europe-west1) on the Gemini Enterprise Agent Platform; per Google, ML processing takes place within the EU. Opus 4.6 introduced Agent Teams and scores 80.8% on SWE-bench Verified per Anthropic, Sonnet 4.6 79.6%.
- Claude Haiku 4.5 (
claude-haiku-4-5, USD 1 / 5, 200K context): Predecessor of Haiku 5.5, still the choice for in-region hosting of small workloads. - Claude Fable 5.1 and Mythos 5.1 (
claude-fable-5-1,claude-mythos-5-1, 1 September 2026, USD 10 / 50): Superseded by Fable 5.2 and Mythos 5.2 respectively since 9 November 2026, still active per Anthropic. Fable 5.1 scored 55.8% on Terminal-Bench 4.0. - Claude Fable 5 (
claude-fable-5): Listed as legacy since Fable 5.1. - Discontinued: Claude Opus 4.1 (5 August 2026), Claude Opus 4 and Sonnet 4 (15 June 2026) and the Claude 3 models. All dates are listed in the AI model archive.
Anthropic: the company behind Claude
Anthropic is a public benefit corporation (PBC) headquartered in San Francisco, California, and was founded in early 2021. Its seven co-founders include Dario Amodei (CEO) and Daniela Amodei (President and Chair of the Board) as well as Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Chris Olah. As a PBC, Anthropic is committed to a public benefit purpose set out in its charter alongside profit. The Long-Term Benefit Trust, whose trustees per Anthropic are independent of management and investors, determines the majority of the board. On 26 October 2026 Anthropic filed a registration statement on Form S-1 with the US Securities and Exchange Commission for a proposed initial public offering; per Anthropic, its Class A common stock is to be listed on Nasdaq under the symbol “ANT”. The number of shares and the price range have not yet been determined.
Important for companies in Germany, Austria and Switzerland: per the Commercial Terms, the contracting party for customers based in the EEA, Switzerland or the UK is Anthropic Ireland, Limited in Dublin, and Irish law applies. For these regions it is also the data protection controller for consumer accounts. Per the Trust Center, the group includes an Anthropic Germany GmbH, an Anthropic France SAS and an Anthropic Switzerland GmbH, among others. In Europe, Anthropic has offices in London, Dublin, Zurich, Paris and Munich; an office in Milan has been announced. Per Anthropic, AWS is its primary cloud and training partner; Amazon and Google are among its investors.
Under the EU AI Act, Anthropic Ireland, Limited has signed the Code of Practice for general-purpose AI models (GPAI Code of Practice). Per the Trust Center, Anthropic holds a SOC 2 Type 2 report and certifications to ISO 27001 and ISO 42001, among others.
What sets Claude apart for enterprises
Constitutional AI: Anthropic anchors ethical principles directly in training. Instead of relying solely on human feedback, the model learns from defined constitutional principles to avoid harmful outputs. In regulated industries this makes compliance easier.
1M token context: All models from the 4.6 generation onwards process up to one million tokens, without a long-context surcharge except on Haiku 5.5 (higher prices for prompts over 100,000 tokens). Per Anthropic, that is roughly 555,000 words with the current tokenizer (from Opus 4.7 onwards) and around 750,000 words on older models. Entire contract corpora or technical documentation can be processed in a single request. Because large contexts raise the cost per request significantly, we recommend keeping the context only as large as needed.
Agentic tools: Computer use (computer_toolset_20260801) and browser use (browser_toolset_20260801) are generally available, as are the Files API, Agent Skills and the Skills API. Managed Agents offer session budgets as a hard cost cap, an effort setting per agent and domain allowlists for web search and web fetch.
Compliance features for Claude Enterprise: With inference hooks (generally available since 20 October 2026), every prompt from claude.ai, Cowork and Claude Code can be sent to your own security server before inference, which approves or blocks it. The Compliance API also returns transcripts of Cowork and Claude Code sessions. The Admin API connects user management to existing identity processes. Important for data protection documentation: the Console and Playground are not covered by zero data retention agreements.
Text watermarking under the EU AI Act: Anthropic marks text generated by Claude with a machine-readable watermark that is imperceptible to humans. Models launched since 2 August 2026 support the marking from launch; older ones are being retrofitted. Per Anthropic it adds no tokens. Anyone using Claude output in publications or proposals should mention this in internal AI policies.
Pricing
| Model | API ID | Input / output (USD per 1M tokens) | Cache read | Context |
|---|---|---|---|---|
| Claude Opus 5.5 | claude-opus-5-5 | 4 / 20 | 0.20 | 1M (128K output) |
| Claude Sonnet 5.5 | claude-sonnet-5-5 | 2 / 10 | 0.10 | 1M (128K output) |
| Claude Haiku 5.5 | claude-haiku-5-5 | 0.10 / 0.50 up to 100,000 prompt tokens, above that 0.50 / 2.50 | 0.01 / 0.05 | 1M (128K output) |
| Claude Fable 5.2 | claude-fable-5-2 | 8 / 40 | 0.20 | 1M (128K output) |
| Claude Mythos 5.2 (after vetting only) | claude-mythos-5-2 | 8 / 40 | 0.20 | 1M (128K output) |
| Claude Fable 5.1 | claude-fable-5-1 | 10 / 50 | 0.25 | 1M (128K output) |
| Claude Fable 5 (legacy) | claude-fable-5 | 10 / 50 | 1.00 | 1M (128K output) |
| Claude Opus 5 | claude-opus-5 | 5 / 25 | 0.50 | 1M |
| Claude Sonnet 5 | claude-sonnet-5 | 2 / 10 | 0.20 | 1M |
| Claude Sonnet 4.6 | claude-sonnet-4-6 | 3 / 15 | 0.30 | 1M |
| Claude Haiku 4.5 | claude-haiku-4-5 | 1 / 5 | 0.10 | 200K |
Anthropic API prices per Anthropic’s pricing page. AWS Bedrock and Google Cloud bill their own prices; Google’s EU multi-region endpoint carries a 10% premium over global.
The Batch API costs 50% of the standard price. Cache writes cost 1.25× the input price (5-minute TTL) or 2× (1-hour TTL). Managed Agents are billed at USD 0.08 per session hour on top. The fast mode costs USD 16 / 80 on Fable 5.2, USD 8 / 40 on Opus 5.5 and USD 10 / 50 on Opus 5 and Opus 4.8 (still a research preview). Since Opus 4.7, a new tokenizer splits the same text into around 30% more tokens; budget planning should therefore measure the cost per task, not the token price.
For production deployments with many users, a model strategy by task pays off:
| Task | Recommended model | Rationale |
|---|---|---|
| Classification, extraction, routing, subagents | Claude Haiku 5.5 | From USD 0.10 / 0.50, 1M context; Haiku 4.5 for in-region hosting |
| Document analysis, support chatbot, everyday coding | Claude Sonnet 5.5 | USD 2 / 10, cache read USD 0.10; Sonnet 5 for in-region hosting on Bedrock |
| Complex agentic work, software engineering, long-horizon agents | Claude Opus 5.5 | Close to Fable 5.1, ZDR-capable; Opus 5 for in-region hosting on Bedrock |
| Hardest reasoning and agentic tasks when Opus 5.5 is not enough | Claude Fable 5.2 | USD 8 / 40, covered model with 30-day retention; in the EU via the Agent Platform |
| Regional EU endpoint instead of multi-region | Claude Sonnet 4.6 / Opus 4.6 | Endpoint in Belgium (europe-west1) on the Gemini Enterprise Agent Platform, ML processing in the EU per Google |
With CompanyGPT, administrators configure the right model per use case; compared with an Opus-only deployment this typically saves 60 to 80% of the cost.
EU availability and GDPR
For a GDPR assessment, what matters most is where processing happens and how long data is retained. The Anthropic API itself offers no EU data residency: the inference_geo parameter only knows "us" and "global", and the only storage location (workspace geo) available is "us". The same applies to the two routes via the hyperscalers that Anthropic operates itself: Claude Platform on AWS (billed via AWS Marketplace) and Claude Models on Google Cloud (Anthropic managed), available since 22 October 2026. For EU workloads we therefore recommend consuming Claude via AWS Bedrock or the Gemini Enterprise Agent Platform (now called “Google managed” by Anthropic), where the cloud provider’s data processing terms and data processing agreements apply.
Where are Claude’s servers?
| Delivery route | Who operates inference? | Data processing per provider |
|---|---|---|
| claude.ai, Claude Team, Claude Enterprise | Anthropic | Routing to countries in the US, Europe, Asia and Australia possible; data is stored in the US. “US-only inference” option only for usage-based Enterprise plans, no EU option |
| Claude API (Claude Platform) | Anthropic | inference_geo global or us, storage in the US |
| Claude Platform on AWS | Anthropic | The workspace’s AWS region does not determine where inference takes place |
| Claude Models on Google Cloud (Anthropic managed) | Anthropic | Location global, workspace geo us |
| AWS Bedrock | AWS | EU geo profile with routing within the EU; in-region Ireland and Stockholm for selected models; London (UK) for Opus 5.5, Sonnet 5, Opus 4.6 and Sonnet 4.6 |
| Gemini Enterprise Agent Platform (Google managed) | Google Cloud | Europe multi-region with ML processing in the EU; for older models a regional endpoint in Belgium (europe-west1), ML processing likewise in the EU |
| Microsoft Foundry | Microsoft or Anthropic | Global Standard or US Data Zone, no EU data zone |
Data protection with claude.ai, Team and Enterprise
- Consumer accounts (Free, Pro, Max): Since 28 August 2025, users decide themselves whether their chats are used for training. If training is allowed, Anthropic retains de-identified data for up to five years; otherwise deleted chats are removed within 30 days. Consumer accounts are therefore unsuitable for business data.
- Team and Enterprise: Per the Commercial Terms, Anthropic does not train on customer content; only explicitly submitted feedback is retained for up to five years, and owners can turn off the feedback function. Without a custom setting, Enterprise retains content indefinitely; custom retention periods from 30 days are only possible on the Enterprise plan. The Team and Enterprise interfaces are not designed for zero data retention (exception: Claude Code on the Enterprise plan).
- Claude API: Per Anthropic, inputs and outputs are deleted in the backend within 30 days unless an exception applies (Files API, zero data retention, flagged policy violations up to two years).
Contracting party, DPA and standard contractual clauses
The contracting party and, for consumer accounts, the controller for users in the EEA is Anthropic Ireland, Limited. For business customers the Data Processing Addendum (effective 24 February 2025) applies: the customer is the controller, Anthropic the processor. The DPA incorporates the EU standard contractual clauses (Modules 2 and 3) and provides for notification of security incidents within 48 hours, a 15-day objection period for new sub-processors and support with transfer impact assessments. Per the Trust Center, the infrastructure sub-processors are Google Cloud, AWS and Microsoft Azure.
When consuming Claude via AWS Bedrock or the Gemini Enterprise Agent Platform, you conclude the data processing agreement with the cloud provider. On Bedrock, per AWS, model providers have no access to prompts and responses; these are only stored for the covered models Fable 5, Fable 5.1 and Fable 5.2 (30 days). For Fable 5.2, per the model card, AWS additionally requires the aws_review data retention mode.
Delivery routes per model
| Model | AWS Bedrock (EU) | Google Cloud, Agent Platform (EU) | Microsoft Foundry |
|---|---|---|---|
| Opus 5.5 | EU geo profile, no in-region in the EU (in-region London, UK) | EU multi-region | Global Standard, US Data Zone |
| Sonnet 5.5 | EU geo profile, no in-region | EU multi-region | Global Standard, US Data Zone |
| Haiku 5.5 | EU geo profile, no in-region | EU multi-region | Global Standard, US Data Zone |
| Fable 5.2 | global./us. only, no EU profile | EU multi-region (GA from day one) | Global Standard (Hosted on Anthropic only) |
| Fable 5.1, Fable 5 | global./us. only, no EU profile | EU multi-region | Preview, Global Standard |
| Opus 5, Sonnet 5, Opus 4.8 | In-region Ireland/Stockholm (bedrock-mantle), EU geo profile | EU multi-region | Global Standard, US Data Zone |
| Opus 4.7 | In-region Ireland/Stockholm (bedrock-mantle), EU geo profile | EU multi-region | Global Standard |
| Haiku 4.5 | In-region Ireland/Stockholm (bedrock-mantle), EU geo profile | Regional endpoint Belgium (ML processing in the EU) | Global Standard, US Data Zone |
| Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5 | EU geo profile only | Regional endpoint Belgium (ML processing in the EU) | Global Standard |
AWS Bedrock: Frankfurt is not an in-region location for any Claude model, only a source region for the EU geo inference profile, whose routing stays within EU geography. True in-region hosting in Ireland (eu-west-1) and Stockholm (eu-north-1) exists only for Opus 5, Sonnet 5, Opus 4.8, Opus 4.7 and Haiku 4.5, and only via the bedrock-mantle endpoint (Anthropic Messages API without Converse, Guardrails, Knowledge Base and Agents). Opus 5.5, Sonnet 5.5 and Haiku 5.5 run in the EU via the EU geo profile on the standard bedrock-runtime endpoint, i.e. with all Bedrock features. Per AWS, Bedrock does not store prompts and responses for any Claude model except the Fable models (5, 5.1 and 5.2). London (eu-west-2) is in the UK, not the EU; per the model cards, Opus 5.5 and Sonnet 5 (since 30 September 2026) as well as Opus 4.6 and Sonnet 4.6 run in-region there on bedrock-runtime. Important for migration: the legacy integration with ARN-versioned model IDs (InvokeModel/Converse) does not support API version 2026-10-21; anyone working through it should pin the SDK version.
Gemini Enterprise Agent Platform: The 5-generation models as well as Opus 4.8 and 4.7 run via the EU multi-region endpoint (aiplatform.eu.rep.googleapis.com) with ML processing in the EU, without a regional endpoint. Opus 4.5, Opus 4.6, Sonnet 4.5, Sonnet 4.6 and Haiku 4.5 instead offer a regional endpoint in europe-west1 (Belgium), not Frankfurt. Per Google’s model pages, ML processing takes place within the Europe multi-region there too; Google does not document a fixed processing location in Belgium. For Fable 5.1, the “Advanced AI Safety Addendum” must be accepted once in the Model Garden. The EU multi-region endpoint does not include the UK and Switzerland. Not to be confused with the new route “Claude Models on Google Cloud (Anthropic managed)”: there Anthropic operates the API itself, and there is no EU data residency.
Microsoft Foundry: Claude is generally available in the Foundry model catalogue, but without an EU data zone deployment. Global Standard (including Sweden Central) is offered, plus a US Data Zone for selected models. Processing is therefore not restricted to the EU, and Anthropic remains an independent data processor. With the “hosted on Azure” variant, per Anthropic, prompts and responses remain in Azure. For strict EU residency we recommend Bedrock or the Agent Platform.
Zero data retention and covered models
- Zero data retention (ZDR) is not the default. ZDR has to be requested from Anthropic per organization. It covers the Messages API, the token counting API and Claude Code with API keys of a commercial organization; the Console, Playground, Managed Agents and code execution are not covered.
- Opus 5.5, Sonnet 5.5, Haiku 5.5 and Opus 5 are not covered models. The regular ZDR rules apply to them.
- Fable 5.2, Mythos 5.2, Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 are covered models. Prompts and outputs are retained for 30 days everywhere, on Bedrock and Google Cloud by the respective cloud provider. ZDR organizations can enable 30-day retention per workspace; without that setting the API rejects requests with
400 invalid_request_error. For processing special categories of personal data or professional secrets this is usually an exclusion criterion.
Recommended EU setups
- Opus 5.5 via the EU multi-region of the Gemini Enterprise Agent Platform or the EU geo profile on AWS Bedrock: per Anthropic the starting point for most workloads, ZDR-capable.
- Opus 5 on AWS Bedrock in Ireland or Stockholm (bedrock-mantle) when a fixed EU location on Bedrock is mandatory.
- Sonnet 5.5 via the EU multi-region or the EU geo profile as the cost-efficient default; Sonnet 5 for in-region hosting in Ireland or Stockholm.
- Haiku 5.5 via the EU multi-region or the EU geo profile for high volume; Haiku 4.5 for in-region hosting on Bedrock in Ireland or Stockholm; on Google Cloud with a regional endpoint in Belgium (ML processing in the EU per Google, can be retired from 7 December 2026).
- Sonnet 4.6 or Opus 4.6 via the regional endpoint of the Agent Platform in Belgium (europe-west1) when a regional endpoint is required instead of the EU multi-region; per Google, ML processing takes place within the EU there too, not at a fixed location in Belgium.
- Fable 5.2 via the EU multi-region of the Agent Platform (GA from day one) when the 30-day retention is documented as acceptable; on Bedrock, Fable 5.2 is only available globally or in the US.
Getting Claude through innFactory
innFactory GmbH from Rosenheim is a cloud reseller for AWS, Microsoft Azure, Google Cloud and STACKIT. For Claude this means: you obtain AWS Bedrock, the Gemini Enterprise Agent Platform or Microsoft Foundry directly through innFactory, by invoice or direct debit instead of credit card, with all cloud providers on one invoice and a personal contact for EU regions and model selection. On this basis, the innFactory AI Gateway, as a component in your own tenant, controls budgets, hard limits, SSO and cost allocation per team and project, including for Claude Code and Cowork. CompanyGPT provides Claude to all employees in a GDPR-compliant interface.
Migration and breaking changes
A plain model ID swap is usually not enough when moving to the current models. The key changes per Anthropic:
- Thinking control: On Opus 5.5, Fable 5.2 and Fable 5.1 adaptive thinking is always on and
thinking: disabledis rejected. On Sonnet 5.5,between_toolsreplacesdisabled(only up to effort high). On Haiku 5.5, thinking can be turned off up to effort high. Manual budgets (budget_tokens) are rejected throughout. - Forced tool choice:
tool_choiceof typeanyortoolis rejected with400on Opus 5.5, Sonnet 5.5, Fable 5.2 and Fable 5.1. Alternatives arestrict: trueor structured outputs. - Thinking blocks: They are bound to the model and the conversation. Conversations that send thinking blocks back must stay append-only; for accounts created on or after 31 August 2026 this is enforced, and on Fable 5.2 and Mythos 5.2 for all accounts. Only Fable 5.2 and Mythos 5.2 read Fable 5.2’s thinking blocks; when switching to another model, the reasoning is lost.
- Computer use: On the Claude API and Google Cloud the 5.5 models only accept
computer_toolset_20260801, and Fable 5.2 rejectscomputer_20251124on all platforms except Bedrock; on Bedrock the older tool still works. - Effort calibration: The default effort is medium on Opus 5.5 and Haiku 5.5 and high on Sonnet 5.5; the levels have been recalibrated. Text between tool calls comes back as a thinking block and is empty at the default display setting.
- Advisor tool (beta): The advisor must be at least as capable as the executor model. Fable 5.2 only accepts Fable 5.2 or Mythos 5.2 as an advisor; Fable 5.1 rejects Mythos 5, Fable 5, Opus 5.5 and Opus 5; Sonnet 5.5 rejects Opus 4.8, Opus 4.7, Sonnet 5 and Haiku 5.5.
- Sampling and prefill: With API version 2026-10-21, Anthropic has removed
temperature,top_pandtop_kfor all models (error 400); older API versions reject non-default values from Claude 4.7 onwards. Assistant prefill has been rejected since Opus 4.7. This also affects the move from Haiku 4.5 to Haiku 5.5. The current SDKs send 2026-10-21 by default. - Tokenizer: Haiku 5.5 uses the newer tokenizer and counts the same text as around 30% more tokens than Haiku 4.5; Sonnet 5.5 uses the same tokenizer as Sonnet 5.
Anthropic provides migration guides for Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1 and Fable 5.2. We recommend configuring model IDs and request parameters centrally so that a switch remains a configuration change.
Model lifecycle
| Model | Status per Anthropic | Retirement on the Claude API | Successor |
|---|---|---|---|
| Claude Sonnet 4.5 | Deprecated since 30 September 2026 | 30 November 2026 (Bedrock: EOL 8 April 2027) | Sonnet 5.5 |
| Claude Opus 4.1 | Discontinued | 5 August 2026 | Opus 4.8 |
| Claude Opus 4 and Sonnet 4 | Discontinued | 15 June 2026 | Sonnet 5.5 |
| Claude Haiku 3 | Discontinued | 20 April 2026 | – |
| Claude Sonnet 3.7 and Haiku 3.5 | Discontinued | 19 February 2026 | – |
| Claude Fable 5 | Legacy | not before 9 June 2027 | Fable 5.2 |
| Claude Opus 5 and Sonnet 5 | Legacy, still active | not before 24 July and 30 June 2027 respectively (Google Cloud: not before 24 January 2027 and 24 December 2026) | Opus 5.5 and Sonnet 5.5 |
| Claude Opus 4.8, 4.7 and 4.6, Sonnet 4.6 | Legacy, still active | not before 28 May, 16 April, 5 February and 17 February 2027 respectively | Opus 5.5 and Sonnet 5.5 |
| Claude Opus 4.5 | Legacy, still active | not before 24 November 2026 | – |
| Claude Haiku 4.5 | Legacy, no deprecation | not yet announced (Google Cloud: not before 7 December 2026) | Haiku 5.5 |
| Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 | Active | not before 22 September, 28 September and 7 October 2027 respectively | – |
| Claude Fable 5.1, Mythos 5.1 | Active, superseded | not before 1 September 2027 | Fable 5.2 and Mythos 5.2 |
| Claude Fable 5.2, Mythos 5.2 | Active | not before 9 November 2027 | – |
Per its own documentation, Anthropic announces retirements at least 60 days in advance. AWS Bedrock and Google Cloud set their own dates: the Bedrock model cards provide for a legacy period of at least six months before EOL, while Google lists earlier minimum dates than Anthropic for some models, such as Sonnet 5 and Haiku 4.5. A cross-vendor overview is in the AI model archive.
How quickly availability can change was shown in June 2026: Fable 5 and Mythos 5 were deactivated worldwide three days after launch by order of the US government and only released again from 1 July 2026. For enterprises this argues for a provider- and model-agnostic setup. Background is in our blog post on GPT-5.6 and Claude Fable 5.
Integration with CompanyGPT
Claude models are integrated into CompanyGPT and can be operated GDPR-compliantly inside your infrastructure. CompanyGPT enables deployment through your own cloud environment (AWS, Google Cloud, or Microsoft Azure) without forwarding data to third parties. Model routing helps optimise cost: simple requests go to a Haiku model, complex ones to a Sonnet model, and demanding reasoning and agentic workloads to an Opus model. This flexibility protects against dependence on a single model: if a model is pulled at short notice, CompanyGPT can switch to another Claude model or another provider.
Typical deployment scenarios include integration into existing knowledge management systems, connection to ERP systems or provisioning as an internal chatbot for employees. With CompanyGPT you retain full control over access rights, audit logs and data flows, essential for industries with strict compliance requirements such as banking, insurance or healthcare.
Our recommendation
For the most demanding GDPR-sensitive workloads: Claude Opus 5.5 (claude-opus-5-5). Per Anthropic on par with Fable 5.1 on most work, cheaper than Opus 5 at USD 4 / 20 per 1M tokens and without mandatory 30-day retention. In the EU via the Europe multi-region of the Gemini Enterprise Agent Platform or the EU geo profile on AWS Bedrock. Recommended for software engineering, long-horizon agents, multi-agent workflows and demanding reasoning. Claude Opus 5 (claude-opus-5) remains the choice if an in-region location on Bedrock in Ireland or Stockholm is mandatory. Claude Fable 5.2 (claude-fable-5-2, USD 8 / 40) pays off for the most demanding agentic tasks where Opus 5.5 is not sufficient even at higher effort, and when the 30-day retention is documented as acceptable; in the EU via the Agent Platform.
For most business applications: Claude Sonnet 5.5 (claude-sonnet-5-5). At the price of Sonnet 5 (USD 2 / 10 per 1M tokens, cache read USD 0.10), per Anthropic more than 30% faster and up to 30% cheaper per task, without mandatory 30-day retention. In the EU via the Europe multi-region of the Agent Platform or the EU geo profile on Bedrock. Claude Sonnet 5 remains the choice for in-region hosting on Bedrock in Ireland or Stockholm, Claude Sonnet 4.6 when a regional endpoint in Belgium (europe-west1) is required; per Google, ML processing takes place within the EU there.
For cost-optimised applications: Claude Haiku 5.5 (claude-haiku-5-5). For high volume and less complex tasks such as categorisation, extraction, routing, summaries or subagents: from USD 0.10 / 0.50 per 1M tokens for prompts up to 100,000 tokens, with a 1M token context and without mandatory 30-day retention. Claude Haiku 4.5 remains the choice if a fixed EU location is mandatory: in-region on Bedrock in Ireland or Stockholm. On Google Cloud there is only a regional endpoint in Belgium with ML processing within the EU multi-region, and the model can be retired there from 7 December 2026.
For procurement in Germany: Claude via AWS Bedrock (EU geo profile) or the Europe multi-region of the Gemini Enterprise Agent Platform, billed through innFactory GmbH as a cloud reseller, with the innFactory AI Gateway for budgets and cost control and CompanyGPT as the interface. The direct Anthropic API, claude.ai and the Anthropic-managed routes are not suitable when EU data residency is required.
Before any model switch, a test with real prompts pays off, because request parameters, effort calibration and token counting differ between generations. As innFactory AI Consulting we support you in choosing the right model, GDPR-compliant integration and cost optimisation. Contact us for a no-obligation initial consultation.