Anthropic Claude from a Swiss fiduciary perspective: residency, pricing, compliance

What is Anthropic Claude?

Anthropic is a U.S. AI safety lab based in San Francisco, founded in 2021 by former OpenAI staff. The Claude model family is known for (a) cautious, well-documented answers, (b) very long context windows, and (c) strong reasoning. Main investors are Amazon and Google – both with multi-billion investments, neither holding a majority stake; Amazon additionally with deep AWS Bedrock integration.

The Claude family is structured in three tiers – a fast, cheap model (Haiku class, good for classification and simple tasks), a balanced generalist (Sonnet class), and a strongest reasoning model (Opus class). Model versions, knowledge cutoffs and per-1M-token prices change continuously; the current names and rates are on Anthropic's official pricing page (order of magnitude as of mid-2026: lower tier around USD 1 / 5, mid tier around USD 3 / 15, top tier around USD 5 / 25 per 1M input/output tokens – without guarantee, verify before use). The models offer very long context windows up into the millions of tokens.

Access runs over three paths: first, api.anthropic.com directly (processing in the U.S., USD billing); second, AWS Bedrock (central for EU fiduciaries: Frankfurt region eu-central-1 as a resident option); third, Google Vertex AI (us-central plus europe-west). Microsoft Foundry with Claude in EU was announced for 2026 but had not yet hit GA as of May 2026.

Why it matters

For Swiss fiduciary offices, Anthropic matters for three reasons. First the profile: Claude answers more cautiously and with clearer source hints than GPT-4o, which leads to the right answer – or to a documented refusal – on legal and tax texts. Second the context window: 1M tokens is enough to fit a complete client annual report plus 5 years of correspondence into a single call. Third the AWS path: AWS Bedrock in Frankfurt eu-central-1 has been a fully resident endpoint since Anthropic's 2024 regionalisation, with AWS's standard DPA including EU SCCs.

The critical points: Anthropic itself is a U.S. company subject to the U.S. CLOUD Act, as is AWS. The "EU Frankfurt" path reduces third-country transfer to the sub-processor level and simplifies the TIA, but does not eliminate it. Anyone who needs absolute EU sovereignty (e.g., defence, strict FINMA reading) stays with Mistral-EU or self-hosted.

The data posture is favourable: Anthropic's Commercial Terms contractually exclude training on customer data (covering Claude for Work, Enterprise, API with DPA). Standard retention is 30 days; Zero Data Retention is available on request. In the consumer tier (Claude.ai Free/Pro), an opt-out toggle for training was introduced in August 2025 that has drawn GDPR criticism – irrelevant for professional data but worth knowing.

How it works

Three contract tiers with different postures. On Claude.ai Free/Pro (the consumer UI), a toggle was introduced in August 2025: default "on" for training, opt-out possible, retention up to 5 years if enabled. This tier is out for professional data. On the API with Commercial Terms (api.anthropic.com plus standard contract), training is contractually excluded by default, retention is 30 days for abuse detection, ZDR is available on request. On Claude Enterprise (Team/Org plans) and via AWS Bedrock/Vertex, the hyperscaler's commercial posture applies.

AWS Bedrock is the most important Swiss path: Claude models run in eu-central-1 (Frankfurt) as a resident endpoint. Ordering via AWS account, authentication via AWS SigV4. Contractually the AWS Customer Agreement plus the AWS Data Processing Addendum apply (current: EU SCCs included). Inference profiles can optionally be set to "EU only" to prevent cross-region routing. Surcharge: roughly 10% versus the global endpoint.

Technically Anthropic has its own API (anthropic-version header, messages.create endpoint), but LiteLLM translates from the OpenAI schema. Tool calling and JSON output are supported. Notable: "thinking content" (reasoning visibility) is off by default on the Opus model in some versions and must be opted in. Prompt caching (up to 90% savings on recurring context) and the Batch API (50% off) are available and cost-relevant for fiduciary RAG.

Claude decision in 6 steps (fiduciary CIO)

1. 01Decide the contract path: AWS Bedrock Frankfurt (EU-resident standard for Swiss fiduciaries) vs. api.anthropic.com directly (U.S., fastest model access) vs. Vertex AI europe-west.
2. 02Obtain the DPA: standardised in the AWS Customer Agreement; for Anthropic direct, request via the Trust Center. EU SCCs are standard in both paths.
3. 03Document the TIA: AWS Frankfurt reduces transfer to the sub-processor level; name the residual U.S. CLOUD Act risk explicitly.
4. 04Model mapping: Haiku model for classification/triage, Sonnet model as the generalist default, Opus model only for complex reasoning cases (several times the cost vs Sonnet).
5. 05Set the "EU-only" inference profile in AWS Bedrock to prevent cross-region routing. Accept the ~10% surcharge.
6. 06Cost controls: enable prompt caching (90% savings on long RAG contexts), use the Batch API for background jobs (50% off), run a monthly AWS Cost Explorer review.

When to use Claude

Claude is the right choice when (a) long documents must be analysed in a single call (annual report, contract, legal pleading), (b) caution and traceable answers matter more than creative originality, (c) AWS Bedrock Frankfurt is an acceptable EU path, and (d) a second model family alongside OpenAI is wanted in Multi-LLM routing (to reduce single-vendor risk).

Concrete fiduciary use-cases: contract review (entire client contract plus Swiss CC/CO excerpts in one prompt), VAT guideline research (ESTV PDFs in full context), annual-report plausibility (balance sheet plus prior years as context), legal first research (lawyers' pleadings with source tagging). The Opus model for complex tax cases, the Sonnet model for the generalist tier, the Haiku model for fast classification.

For Multi-LLM routing, Claude is the counterweight to OpenAI: different architecture, different training data, different weak spots. If you suspect a hallucination from OpenAI, you can send the same prompt to Claude in parallel and compare answers (arena mode). That is not double the cost – it is double the assurance on critical cases.

When not to use

Claude is the wrong choice when (a) the use-case calls for creative originality (marketing copy, slogans, free story generation) – Claude is cautious to the point of dullness; (b) absolute EU sovereignty without U.S. corporate involvement is required (defence, strict FINMA) – even AWS Frankfurt is a sub-processor under the U.S. CLOUD Act; (c) the use-case is cost-sensitive and simple – the Haiku model is cheap but small models like Mistral Small (order of magnitude USD 0.20/0.60 per provider) or self-hosted Llama 3.1 8B is cheaper.

Further caveats: Claude is good at code but Codestral and GPT-4.1 often lead specific coding benchmarks. For vision, Claude is solid but GPT-4o has the broader vision toolchain. For speech-to-text, Anthropic has no own model – Whisper (OpenAI) or Deepgram remain standard.

Critical: anyone using Claude.ai Free/Pro for professional work potentially violates Art. 321 SCC (professional secrecy) – that tier has training by default and 5-year retention. For fiduciary use, ALWAYS go through the Commercial API, Claude for Work, or AWS Bedrock, NEVER through the consumer frontend.

Trade-offs

FAQ

Related topics

Sources

1. Anthropic – Pricing (offizielle Preisliste, Modelle Haiku/Sonnet/Opus 4.x) · 2026-05
2. Anthropic – Data Residency (Regions: AWS Bedrock, Vertex AI, Foundry-Roadmap) · 2026-04
3. AWS Bedrock – Regional Model Availability (eu-central-1 Frankfurt) · 2026-05
4. Anthropic Consumer Terms Update – Training Opt-In/Opt-Out (Aug 2025) · 2025-08