Anthropic Claude

As an AI consultancy from Rosenheim we help companies in Germany, Austria and Switzerland with the GDPR-compliant integration of Anthropic Claude. New since 22 September 2026: With Claude Opus 5.5 (claude-opus-5-5) Anthropic launches the Claude 5.5 family — per Anthropic on par with Fable 5.1 on most work, at USD 4 / 20 per 1M tokens and without mandatory 30-day retention; Sonnet 5.5 and Haiku 5.5 are to follow in the coming weeks. That makes Opus 5.5 our new recommendation for GDPR-sensitive workloads (details in the section below). Since 1 September 2026: Anthropic has released Claude Fable 5.1 (claude-fable-5-1) and the access-restricted Claude Mythos 5.1 (claude-mythos-5-1) — the successors to Fable 5 and Mythos 5 at the unchanged token price of USD 10 / 50 per 1M tokens, but with much cheaper cache reads (USD 0.25). Through CompanyGPT the Opus model Claude Opus 5 (claude-opus-5, 24 July 2026), Claude Sonnet 5 (claude-sonnet-5, 30 June 2026), along with Opus 4.8 and Haiku 4.5, can be operated securely inside your own cloud — via AWS Bedrock (Ireland and Stockholm in-region for Opus 5 and Opus 4.8, per the model card only via the bedrock-mantle endpoint/Anthropic Messages API; from Frankfurt only via the EU cross-region inference profile) or the Gemini Enterprise Agent Platform (formerly Vertex AI); Microsoft Foundry offers Claude as well, but so far without an EU data zone. Note on the Mythos class: Fable 5.1, Fable 5 and Mythos 5.x are “covered models” with mandatory 30-day data retention; zero data retention is available only with explicit authorization by Anthropic. In the EU, Fable 5.1 is available with data residency via the Gemini Enterprise Agent Platform (Europe multi-region, ML processing in the EU); on AWS Bedrock, by contrast, Fable 5.1 and Fable 5 run only via the global. and us. profiles. Because of the mandatory 30-day retention, the Opus class remains our recommendation for workloads that require zero data retention — Opus 5.5 since 22 September 2026, and still Opus 5 for in-region hosting on Bedrock. Direct model access to Mythos 5.1 is limited to Project Glasswing participants; Claude Enterprise customers use the Mythos capabilities indirectly through Claude Security, which has run on Mythos 5.1 since 1 September 2026. This overview reflects the state as of 22 September 2026.

Claude 3P: Anthropic as a multi-cloud model

Since the Sonnet 3 generation Anthropic has pursued a consistent multi-cloud strategy. Claude models are not only delivered via Anthropic’s own API but in parallel through the three large hyperscalers. Anthropic refers to this third-party distribution internally and in its documentation as Claude 3P (Third Party). For enterprise buyers this is usually the more relevant variant because it enables GDPR-compliant use through familiar EU cloud regions.

The three Claude 3P paths are:

• AWS Bedrock, the Claude platform on Amazon. Anthropic has been a strategic AWS partner since 2023 and Claude models are a first-class citizen of Bedrock. Frankfurt (eu-central-1) is, per the AWS model cards, never an in-region location for any current Claude model — only ever a source region for the EU cross-region inference profile (routing stays inside the EU geography). True in-region hosting exists from Opus 4.7 onward (incl. Opus 4.8, Opus 5, Sonnet 5, Haiku 4.5) exclusively in Ireland (eu-west-1) and Stockholm (eu-north-1) — and only via the newer bedrock-mantle endpoint (Anthropic Messages API, without Guardrails/Knowledge Base/Agents/Flows). The standard bedrock-runtime endpoint (Converse/InvokeModel, every Bedrock-native feature) offers only the EU geo profile or global there, as in Frankfurt; for Opus 4.5/4.6 and Sonnet 4.5/4.6 there is currently no in-region hosting in the EU on Bedrock at all.
• Gemini Enterprise Agent Platform (formerly Vertex AI), the Google Cloud platform for models and agents. Google renamed Vertex AI to Gemini Enterprise Agent Platform at Cloud Next 2026 on 22 April 2026, all previous Vertex AI capabilities are included. Claude runs via the Model Garden; per Google’s model pages, Opus 4.5, Opus 4.6, Sonnet 4.5, Sonnet 4.6 and Haiku 4.5 offer a single-region option in europe-west1 (Belgium) — not Frankfurt. Fable 5.1, Fable 5, Opus 5, Opus 4.7, Opus 4.8 and Sonnet 5 run there only via the EU multi-region endpoint or global, with no single-region option.
• Microsoft Foundry (formerly Azure AI Foundry), the Microsoft platform for models and agents. Anthropic models have been GA in the model catalogue as Claude 3P since early July 2026. Important caveat: Claude currently runs in Foundry only as a Global Standard deployment (hosted in Sweden Central and East US 2, among others), plus Data Zone Standard (US) for the Hosted-on-Azure variants of Opus 5, Opus 4.8 and Sonnet 5 — there is no EU data zone deployment, so processing is not restricted to the EU and Anthropic remains the independent data processor. Microsoft lists “Foundry in Europe” as coming in 2026.

For GDPR-compliant deployments in the DACH region the de-facto choice has settled on consuming Claude through AWS Bedrock (EU geo profile, or in-region Ireland/Stockholm via bedrock-mantle for Opus 4.7+/Sonnet 5/Haiku 4.5) or the Agent Platform (single-region Belgium or EU multi-region). Given the missing EU data zone, Microsoft Foundry is currently only an option for Claude when strict EU residency is not required.

Key strengths

Constitutional AI

Claude is built on Constitutional AI, a safety approach that anchors ethical principles directly in the training process. Instead of relying exclusively on human feedback the model learns to avoid harmful or inappropriate output against a defined set of constitutional principles. For enterprises this means a lower risk profile when interacting with sensitive data and higher predictability in production. In regulated sectors such as finance, healthcare and public administration this approach helps demonstrate compliance.

Extended context (1M token GA)

Claude offers a context window of one million tokens, generally available (GA) at no premium for Opus 5, Sonnet 5, Sonnet 4.6, Opus 4.6, Opus 4.7 and Opus 4.8. That corresponds to roughly 700,000 words or several hundred pages of documentation. For enterprise use this opens up new options. Entire contract corpora, technical documentation or large project archives can be processed in a single request. API cost note: A 1M token context increases the cost per request significantly. For cost-optimised setups we recommend keeping the context only as large as needed and using Haiku 4.5 for simpler tasks.

Claude Opus 5.5: start of the Claude 5.5 family (22 September 2026)

On 22 September 2026 Anthropic released Claude Opus 5.5 (claude-opus-5-5), the first model of the new Claude 5.5 family; Sonnet 5.5 and Haiku 5.5 are to follow in the coming weeks according to the announcement. Anthropic positions Opus 5.5 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. It is the first release since Anthropic publicly called for pacing the frontier, and it was tested before release by external evaluators (including METR).

Performance (per Anthropic, adaptive thinking at max effort): Terminal-Bench 4.0 66.4% (Fable 5.1: 55.8%, Opus 5: 52.3%, GPT-6 Astra: 57.9%), FrontierCode v1.1 54.4% (50.3% / 48.0%), CursorBench 4.0 57.8%, GDPval-AA v2.1 1846 Elo (1735 / 1708), AutomationBench 40.0% (measured by Zapier, without fallback models), Humanity’s Last Exam 67.7% with tools, Terminal-Bench-Science 58.7%, OSWorld 2.0 81.8% (partial) and Chartography 89.0%. Anthropic itself notes that at this level benchmark margins have become a less reliable guide to real-world differences and that the gap to Fable 5.1 is narrower in practice. Concrete examples from the announcement: a 200,000-line codebase was audited and fixed in under three hours (Opus 5: over 20 hours with 2.5× the tokens); translating HAProxy from C to Rust took 9.5 hours versus 12 hours for Fable 5.1, at 51% lower cost.

Pricing: USD 4 per 1M input tokens, USD 20 per 1M output tokens — 20% below Opus 5 (USD 5 / 25). Cache reads cost USD 0.20 (Opus 5: USD 0.50), cache writes USD 5 (5-minute TTL) or USD 8 (1-hour TTL), the Batch API USD 2 / 10. Because the model also uses fewer tokens per task and generates output more than 30% faster, Anthropic puts the saving on typical workloads at around 40%. A fast mode (research preview, up to 2.5× faster) is available only via the Claude API and in Claude Code and costs USD 8 / 40. Context window 1M tokens, maximum output 128K tokens (300K via the Batch API beta), knowledge cutoff June 2026, retirement not before 22 September 2027.

API changes when migrating from Opus 5: Four breaking changes. First, adaptive thinking is always on — thinking: disabled and manual budgets are rejected with 400 invalid_request_error; depth is controlled by the effort parameter, whose default is now medium instead of high. Second, forced tool choice (tool_choice of type any or tool) is rejected, as already on Fable 5.1; the alternatives are strict: true or structured outputs. Third, thinking blocks are bound to the model and the conversation: Opus 5.5 reads blocks from Opus 5 and earlier Opus, Sonnet and Haiku models, but not from Fable or Mythos models; for organizations created on or after 31 August 2026 the binding to an unchanged prefix is enforced. Fourth, the Claude API and Google Cloud no longer accept the older computer use tool computer_20251124, only the computer_toolset_20260801 toolset (the old tool keeps working on Bedrock). Without an error, but with an effect: text between tool calls comes back as a thinking block that is empty at the default display setting — applications that stream such progress notes must set thinking.display. Anthropic provides a migration guide.

Safety and safeguards: Per Anthropic, Opus 5.5 achieves the best scores of any model tested to date on its automated behavioural audit, is much less likely to take hard-to-reverse actions and is more resistant to prompt injection than Opus 5. Because the model is comparable to Mythos 5.1 in biology and cybersecurity, it ships with safeguards similar to Fable 5.1: a biology safety classifier joins the cybersecurity one, and requests that push the model to reproduce its internal reasoning can be declined with the reasoning_extraction category. Verification programmes exist for biology and cyber use (Life Sciences Verification Program, Cyber Verification Program). Declined requests return stop_reason: "refusal"; server-side fallback (fallbacks: "default", beta) is available.

EU availability and data protection: The key difference to Fable 5.1: per Anthropic’s documentation Opus 5.5 is not a covered model, so zero data retention is possible. In the EU it is available from day one via the Gemini Enterprise Agent Platform with data residency — per Google’s model page GA in the Europe multi-region with ML processing in the EU. On AWS Bedrock the EU cross-region inference profile eu.anthropic.claude-opus-5-5 is available (source regions include Frankfurt, Zurich, Stockholm, Ireland, Paris, Milan, Madrid), but per the model card there is no in-region endpoint in Europe — the bedrock-mantle endpoint is in-region for Opus 5.5 only in us-east-1, Melbourne and GovCloud. Anyone who needs Ireland or Stockholm as a fixed location stays on Opus 5 for now. Microsoft Foundry offers Opus 5.5 as GA (Hosted on Azure and Hosted on Anthropic), Global Standard and US Data Zone, but still no EU data zone. Our assessment: For GDPR-sensitive workloads Opus 5.5 is the new recommendation — near-frontier capability, ZDR-capable, cheaper than Opus 5, available in the EU via Google Cloud with data residency or via the EU geo profile on Bedrock.

Claude Fable 5.1 & Mythos 5.1: new top of the Mythos class (1 September 2026)

On 1 September 2026 Anthropic released Claude Fable 5.1 (claude-fable-5-1) and Claude Mythos 5.1 (claude-mythos-5-1). According to Anthropic’s documentation Fable 5.1 is the current “latest” model and available to all customers via the Claude API and the partner platforms; Fable 5 has been listed as “legacy” since then (retirement not before 9 June 2027), and Anthropic recommends migrating. Mythos 5.1 is technically the same model with more permissive safeguards and remains reserved for Project Glasswing participants — per Anthropic currently exclusively a set of US organizations; an expansion to further domestic and international partners is being coordinated with the US government.

Performance (figures from Anthropic): The jump over Fable 5 shows mainly on agentic tasks: Terminal-Bench 4.0 55.8% (Fable 5: 42.0%, Opus 5: 52.3%, Mythos 5.1: 60.9%), Terminal-Bench-Science 52.6% (24.7%), AutomationBench 31.4% (17.1%), GDPval-AA v2 1853 Elo (1723), OSWorld 2.0 77.9% partial (72.9%), Humanity’s Last Exam 60.9% without and 65.0% with tools (57.8% / 63.8%) and CursorBench 3.2.0 73.4% (70.5%). Multilingual performance is on par with Fable 5 according to Anthropic. The knowledge cutoff is June 2026, the context window 1M tokens (at the standard price across the entire window), the maximum output 128K tokens.

Pricing: USD 10 per 1M input tokens, USD 50 per 1M output tokens — unchanged from Fable 5. New is the cache read price of USD 0.25 per 1M tokens (Fable 5: USD 1; Opus 5: USD 0.50). Anthropic puts the resulting savings at around 25% for typical workloads and up to 45% for heavily agentic workloads. Cache writes cost USD 12.50 (5-minute TTL) or USD 20 (1-hour TTL), the Batch API USD 5 / 25. There is no priority tier for Fable 5.1, Mythos 5.1, Opus 5 and Sonnet 5. The tokenizer is unchanged from Fable 5, so cost comparisons can be made directly.

API changes for integration: Fable 5.1 brings three breaking changes compared with Fable 5. First, forced tool choice (tool_choice of type any or tool) is rejected with 400 invalid_request_error — only auto and none are permitted; anyone who needs a guaranteed call uses strict: true or structured outputs. Second, thinking blocks are bound to the producing model; older models discard them on a fallback. Third, editing earlier turns invalidates the thinking blocks (“preserved thinking”) — for organizations created on or after 31 August 2026 this is already enforced with an error. Adaptive thinking is always on, thinking: disabled is rejected; all five effort levels (low to max) are available. New in beta are per-message effort changes without cache loss, turn-scoped system messages (clear_at) and thinking.display: "updates". As with every model from the 4.6 generation onwards, temperature, top_p, top_k and assistant prefill are gone. Anthropic provides a migration guide.

Safeguards and fallback: The safety mechanisms have been reworked: according to Anthropic the new cyber safeguards block 60% fewer false positives, Fable 5.1 may be used for vulnerability discovery (not exploit development), and the filters intervene 85% less often on elementary biology and medical questions. At the same time the AWS model card notes that the model’s refusal rates are “materially higher” than on previous Claude models. Anthropic names Opus 4.8 and Opus 5 as permitted fallback targets; the server-side fallbacks: "default" (beta) exists only on the Claude API — on Bedrock, Google Cloud and Foundry the fallback has to be implemented client-side.

EU availability and data protection — the decisive point: The EU route with data residency is the Gemini Enterprise Agent Platform (formerly Vertex AI): Google’s model page lists Fable 5.1 as GA in the Europe multi-region (aiplatform.eu.rep.googleapis.com) and the US multi-region as well as on the global endpoint — with ML processing in the EU multi-region, i.e. within EU geography, at the usual 10% premium over global; retirement there is set no sooner than 1 March 2027. A one-time consent to the “Advanced AI Safety Addendum” in the Model Garden is required. On AWS Bedrock Fable 5.1 is open to all customers, but according to the model card only via the global. and us. inference profiles — there is neither an EU geo profile nor an in-region endpoint in Europe; EU regions such as Frankfurt or Stockholm merely act as sources for global cross-region inference, and the regional endpoint of the Claude Platform on AWS exists only in us-east-1. Microsoft Foundry lists Fable 5.1 as a preview, Global Standard only, and Mythos 5.1 as a gated research preview. Regardless of the route, the status as a covered model applies: like Fable 5, Fable 5.1 requires 30-day retention of prompts and outputs on every platform; without explicit authorization by Anthropic, zero data retention is not available. Anthropic has newly announced the Enterprise Frontier Safeguards for this (see the zero data retention section). For processing that requires zero data retention, Opus 5 therefore remains our recommendation; Fable 5.1 is an option for workloads where the 30-day retention is documented as acceptable — in the EU via the Agent Platform, not via Bedrock. Like every Claude model launched since 2 August 2026, Fable 5.1 carries the statistical text watermark (EU AI Act Art. 50) on every platform.

Claude Fable 5 & Mythos 5: withdrawn in June, Fable 5 back since July 2026

On 9 June 2026 Anthropic released Claude Fable 5 (claude-fable-5), the first publicly available Mythos-class model, alongside the access-restricted Claude Mythos 5 (claude-mythos-5, Project Glasswing). After a three-week withdrawal, Fable 5 has been available again since 1 July 2026; Mythos 5 remains access-restricted.

On 12 June 2026 (5:21 PM ET) the US government issued an export-control / national-security directive requiring all access to Fable 5 and Mythos 5 by any foreign national (inside or outside the US, including foreign-national Anthropic employees) to be suspended immediately. Because nationality cannot be verified in real time at the API layer, Anthropic disabled Fable 5 and Mythos 5 for all customers worldwide — just three days after launch.

The background is an escalating dispute between Anthropic and the US government: the Pentagon had demanded unrestricted access to Claude, and Anthropic refused two uses (fully autonomous lethal weapons and mass domestic surveillance of US citizens). The stated trigger was a government claim of a “potential narrow, non-universal jailbreak” of Fable 5. Anthropic says it received only verbal notice and publicly disagreed that this warranted a recall.

Update 26 June 2026 – Mythos 5 partially re-enabled (US only): US Commerce Secretary Howard Lutnick partially lifted the directive and re-enabled Mythos 5 (the variant without a safety classifier) for a defined list of more than 100 vetted US organizations in critical infrastructure (government agencies and private companies). Fable 5 initially remained offline for all users worldwide at that point, and general users, foreign nationals and EU / non-US customers had no access to either Fable 5 or Mythos 5.

Update 30 June / 1 July 2026 – directive fully lifted, Fable 5 back worldwide: On 30 June 2026 the US Commerce Department fully lifted the export restrictions on both models (Anthropic: “Redeploying Fable 5”). Since 1 July 2026, Claude Fable 5 has been globally available again – on Claude.ai, via the Claude API, in Claude Code and Claude Cowork, and per Anthropic’s docs also via Amazon Bedrock, Vertex AI and Microsoft Foundry (on Bedrock, however, without an EU geo profile, see above). Pricing is $10/MTok input and $50/MTok output; a new safety classifier blocks the disputed jailbreak in over 99% of cases according to Anthropic, with blocked requests falling back to an Opus model. The lift comes with conditions (proactive risk detection, cooperation on standards, reporting of malicious use). Mythos 5, by contrast, remains access-restricted – invitation-only via Project Glasswing for vetted organizations. Since 1 September 2026 both models have been superseded by Fable 5.1 and Mythos 5.1; Fable 5 remains available as a legacy model until at least June 2027.

What this means for EU enterprises: Fable 5 has been usable by EU customers since 1 July 2026 – with one weighty compliance caveat: according to Anthropic, prompts and outputs of Mythos-class models are retained for 30 days on every platform – including Bedrock, Vertex and Foundry endpoints in the EU; existing zero-data-retention agreements do not cover Fable 5 traffic. Anthropic’s documentation lists Fable 5 and Mythos 5 as “covered models”: zero data retention is technically unavailable for them — requests from organizations with a ZDR agreement are rejected by the Anthropic API with a 400 invalid_request_error. The 30-day retention therefore cannot be opted out of; it is a precondition for using the models. For processing involving special categories of personal data or professional secrecy obligations, this is usually a knock-out criterion. Since 24 July 2026, Claude Opus 5 (claude-opus-5) is therefore the recommendation for strict GDPR workflows (such as §203 scenarios): per Anthropic it comes close to Fable 5 (and beats it on some benchmarks) — at half the price, without the 30-day retention mandate and with a zero-data-retention option (the default on Bedrock). Fable 5 is an option for workloads where the 30-day retention is acceptable and documented.

Update of 21 August 2026 – Mythos 5 reaches Enterprise customers via Claude Security: Anthropic announced that it is bringing the cybersecurity capabilities of Mythos 5 “to more defenders”. Concretely, as of that day:

• Claude Security runs on Mythos 5. The service scans codebases for vulnerabilities and suggests patches for human review. Per Anthropic it is in public beta for Claude Enterprise customers and billed at standard token rates. Typical use cases are pre-release audits, legacy code review, scheduled regression scanning and working through an existing vulnerability backlog.
• Partner integrations: security product vendors can integrate Mythos 5 into their tools; Anthropic collects expressions of interest through a dedicated form.
• Cyber Verification Program: security teams performing legitimate cybersecurity work can apply through Anthropic support. Approved organizations receive reduced safeguards on Opus and Sonnet models; Anthropic says Mythos-class access is to follow at a later point.
• Defender Advantage Fund: Anthropic is providing USD 35 million in credits for open-source security work (patching live vulnerabilities, automating scanning and patching, and helping projects pursue more ambitious security approaches).

Important context: this does not make Mythos 5 generally orderable through the API. Enterprise customers receive the capabilities mediated through the Claude Security product; direct model access remains gated on approval. For the GDPR assessment, one thing is unchanged: Mythos-class models are “covered models” with no zero-data-retention option. Anyone planning to run Claude Security against production source code should map the 30-day retention, the processing location and the GitHub repository integration in their records of processing activities beforehand.

Only a few days later, OpenAI followed a similar pattern: GPT-5.6 (Sol/Terra/Luna) also launched on 26 June 2026 only as a government-cleared limited preview (see OpenAI GPT). The case underlines why a provider- and model-agnostic, data-sovereign setup matters for EU enterprises. If a single model can be withdrawn overnight, that must not jeopardize business operations. CompanyGPT secures exactly that by flexibly routing requests across Opus 5, Sonnet 5 or models from other providers. Read more in our blog post on GPT-5.6 and Claude Fable 5.

Claude Sonnet 5: new mid-tier flagship (30 June 2026)

With Claude Sonnet 5 (claude-sonnet-5) Anthropic released its most capable Sonnet model to date on 30 June 2026. It supersedes Sonnet 4.6 as the default mid-tier recommendation and, according to Anthropic, significantly closes the gap to the Opus class (Opus 4.8) — at a fraction of the cost. Anthropic highlights, among others, 34.6% on Humanity’s Last Exam (no tools) and 46.8% (with tools), as well as tangible gains in coding, agentic web search (BrowseComp) and computer use (OSWorld-Verified).

Pricing (update of 10 August 2026): The rate communicated at launch as an introductory price — USD 2 per 1M input tokens and USD 10 per 1M output tokens — has been the regular standard price since 10 August 2026. The increase to USD 3 / USD 15 originally announced for 1 September 2026 is cancelled for good. Sonnet 5 therefore sits below the price of Sonnet 4.6 (USD 3 / 15) not just temporarily but as a matter of course — which makes larger rollouts predictable, because there is no longer a cut-off date to plan around.

EU availability: Sonnet 5 is available via AWS Bedrock in-region in Ireland (eu-west-1) and Stockholm (eu-north-1) — per the model card exclusively via the newer bedrock-mantle endpoint (Anthropic Messages API, without Converse/InvokeModel and without Guardrails/Knowledge Base); via the standard bedrock-runtime endpoint only the EU cross-region inference profile or global is available there. The Gemini Enterprise Agent Platform (formerly Vertex AI), announced only as “coming soon” at launch, has since caught up: Sonnet 5 (claude-sonnet-5) runs there with a 1M token context window via the global and the EU multi-region endpoint — routing stays inside EU geography, and no single-region endpoint exists for Sonnet 5. Microsoft Foundry offers Sonnet 5 as well, but so far without an EU data zone (see the Foundry note above). Through CompanyGPT Sonnet 5 can therefore be operated GDPR-compliantly inside your own cloud.

Claude Opus 5: flagship of the Opus class (24 July 2026)

With Claude Opus 5 (claude-opus-5) Anthropic released the first Opus model of the fifth Claude generation on 24 July 2026 — its fourth new model in under two months (after Mythos 5, Fable 5 and Sonnet 5). The positioning is remarkable: per Anthropic, Opus 5 comes close to the frontier intelligence of Fable 5 — at half the price — and even beats Fable 5 on some benchmarks. On Frontier-Bench v0.1 and GDPval-AA (knowledge work) Opus 5 is the new state of the art according to Anthropic, with roughly twice the Frontier-Bench performance of Opus 4.8. On CursorBench 3.2 it lands within 0.5% of Fable 5 at about half the cost per task, on ARC-AGI 3 it scores roughly 3× the next-best model per Anthropic, and on OSWorld 2.0 (computer use) it takes the top score at a third of Fable 5’s cost. In specialist domains Anthropic reports +10.2 percentage points on organic chemistry and +7.7 percentage points on protein tasks versus Opus 4.8. Only on cybersecurity exploitation and biology research does Mythos 5 stay ahead. Anthropic describes Opus 5 as much stronger at verifying its own work and iterating carefully until it actually succeeds; on the automated behavioral audit it posts the lowest misalignment score of recent Claude models at 2.3.

Pricing: USD 5 per 1M input tokens, USD 25 per 1M output tokens — identical to Opus 4.8. The capability jump comes without a price increase. New is an optional fast mode at twice the price that responds roughly 2.5× faster. Performance scales with the effort setting, letting you trade maximum intelligence against token and latency savings per use case. Adaptive thinking is on by default. Opus 5 is the new default model on Claude Max and the strongest model on Claude Pro.

API details for integration: For Opus 5 the 1M token context window is both the default and the maximum, with no long-context surcharge (this applies to every model from the 4.6 generation onwards). Thinking is enabled by default and controlled through the effort ladder low, medium, high, xhigh and max; thinking: disabled is only permitted up to effort high — at xhigh and max the API returns an error (400). The knowledge cutoff is May 2026. Fallback behaviour is controlled through a dedicated fallbacks parameter with a "default" mode.

New platform features: Opus 5 ships mid-conversation tool changes (beta) and Automatic Fallbacks (beta): when a safety classifier triggers, the request is automatically routed to another model and returns a functional response instead of an error. Enterprises also gain access to the Cyber Verification Program. The cybersecurity safeguards remain in place but are expected to engage about 85% less often than on Fable 5, according to Anthropic.

GDPR-relevant — no 30-day retention mandate: Unlike Fable 5, Opus 5 has no data retention requirements for general access; on AWS Bedrock zero data retention is even enabled by default. Opus 5 therefore replaces Opus 4.8 as our recommendation for the most demanding workloads in GDPR-sensitive environments.

EU availability Opus 5 (as of 24 July 2026): On AWS Bedrock Opus 5 is available from day one in-region in eu-west-1 (Ireland) and eu-north-1 (Stockholm) — per the model card only via the newer bedrock-mantle endpoint (Anthropic Messages API, without Converse/InvokeModel and without Bedrock-native features such as Guardrails or Knowledge Base). Via the standard bedrock-runtime endpoint, only the EU cross-region inference profile (eu.anthropic.claude-opus-5) or global is available there, as from Frankfurt (eu-central-1), Zurich (eu-central-2), Paris (eu-west-3), Madrid (eu-south-2), Milan (eu-south-1) and London (eu-west-2) — requests are routed exclusively inside EU geography either way. On the Gemini Enterprise Agent Platform Opus 5 runs via the global and the EU multi-region endpoint (1M token context, no single-region endpoint). Microsoft Foundry offers Opus 5 from day one, but so far without an EU data zone (global standard routing, see above). No in-region endpoint in Frankfurt applies to every Claude version on Bedrock — anyone needing strict single-region hosting at one fixed EU location with the full Bedrock feature set (Converse, Guardrails) will only find that via the Gemini Enterprise Agent Platform in Belgium (europe-west1, available for Opus 4.5/4.6, not for Opus 5).

Claude Opus 4.8: previous flagship (28 May 2026)

Claude Opus 4.8 (claude-opus-4-8) was Anthropic’s most capable model with a zero-data-retention option from late May until 24 July 2026 and remains GA; since 24 July 2026 Claude Opus 5 has been the most capable Claude model with a zero-data-retention option, at identical pricing (see above). The gap to Opus 4.7 is measurable rather than cosmetic: Anthropic reports 69.2% on Terminal-Bench 2 (Agentic Coding, up from 64.3% on Opus 4.7), 57.9% on HCAST with tools (multi-discipline reasoning, up from 54.7%) and a jump to 1890 Elo on Anthropic’s internal knowledge-work evaluation (up from 1753). Anthropic describes the model as having “sharper judgement, more honesty about its progress, and the ability to work independently for longer than its predecessors” — early testers report that Opus 4.8 is more likely to flag uncertainty explicitly instead of confabulating.

The release also ships two new platform features:

• Dynamic Workflow: Opus 4.8 can orchestrate multiple specialised sub-agents in parallel within a single request, moving Anthropic away from the serial single-agent model and competing more directly with multi-agent architectures.
• Effort control panel: Callers can dial how much “effort” (reasoning depth, tool calls, token budget) Claude is allowed to spend on a response. For CompanyGPT deployments this can be pinned per use case, making cost and latency predictable.

Crucially, pricing is identical to Opus 4.7 — USD 5 per 1M input tokens, USD 25 per 1M output tokens. Since Opus 5 (24 July 2026) ships at the same price as well, new projects should start directly on Opus 5; existing 4.7/4.8 deployments usually migrate with a model-ID swap.

EU availability Opus 4.8 (as of 30 May 2026): On AWS Bedrock Opus 4.8 is in-region in eu-west-1 (Ireland) and eu-north-1 (Stockholm) — per the model card only via the bedrock-mantle endpoint (Anthropic Messages API). Via bedrock-runtime (Converse/InvokeModel), only the EU cross-region inference profile (eu.anthropic.claude-opus-4-8) or global is available there, as from Frankfurt (eu-central-1), Zurich (eu-central-2), Paris (eu-west-3), Madrid (eu-south-2), Milan (eu-south-1) and London (eu-west-2); requests are routed inside EU geography. On the Gemini Enterprise Agent Platform Opus 4.8 is reachable only through the EU multi-region endpoint (no single-region); Microsoft Foundry offers Opus 4.8 in Sweden Central from day one. Note: Opus 4.6 is likewise not deployed as an in-region model in Frankfurt on Bedrock (only the EU geo profile/global there) — strict single-region hosting at one fixed EU location with the full Bedrock feature set does not currently exist for the Opus class on Bedrock at all; the Gemini Enterprise Agent Platform in Belgium (europe-west1) is the alternative for Opus 4.6.

Claude Opus 4.7: previous flagship (16 April 2026)

Claude Opus 4.7 was the top-tier model from April through late May 2026 and remains GA. Since Opus 4.8 ships better numbers on coding, tool-augmented reasoning and knowledge work at the same price, 4.7 no longer has a distinct use case for new builds. Existing deployments can usually migrate to 4.8 with a model-ID swap. EU availability: Bedrock, Gemini Enterprise Agent Platform (EU multi-region) and Microsoft Foundry Sweden Central.

Claude Sonnet 4.6: proven predecessor (superseded by Sonnet 5 since June 2026)

With Claude Sonnet 4.6 (claude-sonnet-4-6) Anthropic fundamentally shifted the cost structure for capable AI. The model reaches benchmark scores at Opus level (SWE-bench 79.6%, HumanEval 94.2%) at a significantly lower price point. Sonnet 4.6 is GDPR-compliantly available via AWS Bedrock (EU geo profile from Frankfurt or global only, no in-region hosting) and the Gemini Enterprise Agent Platform — including strict single-region hosting in Belgium (europe-west1, not Frankfurt). Since 30 June 2026 Claude Sonnet 5 has been the new mid-tier flagship (see above); Sonnet 4.6 nevertheless remains a valid choice when strict single-region hosting at one fixed EU location is required — the Agent Platform in Belgium is the fit there, since Sonnet 5 has no single-region endpoint (on Bedrock, Sonnet 5 only offers in-region in Ireland/Stockholm via the bedrock-mantle endpoint).

For mid-sized companies in the DACH region this means enterprise-grade AI performance at cost levels that make larger rollouts economically viable. With 89.1% on MMLU-Pro Sonnet 4.6 demonstrates excellent performance on complex reasoning tasks, ideal for use cases such as automated customer service, technical support, or business process analysis.

Agent Teams (Claude Opus 4.6)

Claude Opus 4.6 (claude-opus-4-6) introduced the Agent Teams feature, where multiple specialised AI agents collaborate in an orchestrated way. One agent handles data analysis, a second produces the report, a third performs quality control. This division of labour enables workflows that overwhelm a single model.

With 80.8% on SWE-bench Verified and 68.8% on ARC-AGI-2 Opus 4.6 shows particular strength on demanding coding and reasoning work. Opus 4.6 is available single-region in the EU through the Gemini Enterprise Agent Platform in Belgium (europe-west1); on AWS Bedrock there is no in-region hosting for it, only the EU geo profile from Frankfurt or global. Cost note: Opus models are markedly more expensive than Sonnet. Use only when Opus-tier capability is genuinely required.

Project Glasswing: Mythos in cybersecurity research

Mythos-class models originate from Project Glasswing, Anthropic’s cybersecurity initiative. The predecessor Mythos Preview discovered thousands of previously unknown zero-day vulnerabilities across every major operating system and web browser, including a 27-year-old vulnerability in OpenBSD and a 16-year-old vulnerability in FFmpeg. The models introduced in June 2026 — Mythos 5 (without safety classifiers, for approved partners only) and Fable 5 (with safety guardrails for production scenarios) — were, however, withdrawn worldwide by order of the US government just three days after launch (see the section “Claude Fable 5 & Mythos 5”). On 30 June 2026 the directive was fully lifted. Since 1 September 2026 Mythos 5.1 is the current Glasswing variant: technically identical to Fable 5.1, but with more permissive safeguards for vetted organizations — per Anthropic currently only a set of US organizations, with access via the Anthropic, AWS or Google Cloud account team. Claude Security has run on Mythos 5.1 since then (previously on Mythos 5 since 21 August 2026) and gives Claude Enterprise customers indirect access to the Mythos capabilities for vulnerability scans. Through the Cyber Verification Program vetted security teams currently receive reduced safeguards on Opus and Sonnet models; Anthropic says access to Mythos-class models through this program will follow “in the near future”. In addition, on 27 August 2026 Anthropic announced a Life Sciences Verification Program whose first participants were enrolled in coordination with the US government.

Note on API cost

Claude models are expensive compared with other providers. For production deployments with many users we recommend a deliberate model strategy. Current token prices can be compared transparently at ai-prices.eu.

With CompanyGPT administrators can configure the right model per use case, typical savings of 60 to 80 percent compared with a pure Opus deployment become achievable.

Anthropic API price overview (as of 22 September 2026)

The discount and surcharge mechanics matter just as much for cost planning: the Batch API costs 50% of the standard price, cache reads are billed at 0.1× the input price — on Fable 5.1 and Mythos 5.1 at only 0.025× (USD 0.25) since 1 September 2026 — and cache writes at 1.25× (5-minute TTL) or 2× (1-hour TTL). The 1M token context is included in the standard price without a long-context surcharge for every model from the 4.6 generation onwards. Managed Agents are billed at USD 0.08 per session hour on top of token cost.

Important for cost comparisons — new tokenizer since Opus 4.7: Since Opus 4.7 Anthropic uses a new tokenizer that splits the same text into roughly 30% more tokens. Comparing per-token prices alone is therefore not meaningful, neither against older Claude models nor against other providers. For budgeting and business cases, always measure cost per task rather than the token price. In migration projects we plan a corresponding buffer whenever a deployment moves from an older to a newer Claude model.

Zero data retention, data retention and data residency

For a GDPR assessment the model choice matters less than the questions of where processing happens and how long data is retained. The picture as of 22 September 2026:

• Zero data retention (ZDR) is not the default. ZDR has to be requested from Anthropic and enabled per organization. Without an explicit agreement, regular retention applies.
• What ZDR covers: the Messages API, the token counting API, and Claude Code (with API keys belonging to a commercial organization).
• What ZDR does not cover: Console and Playground (called “Workbench” until 18 August 2026), Managed Agents (because they are stateful), and code execution. Anyone using these building blocks should reflect that properly in their records of processing activities and technical measures.
• Opus 5.5 is not a covered model. For Opus 5.5, released on 22 September 2026, the regular ZDR rules apply per Anthropic’s documentation — unlike the Mythos class.
• Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 are excluded. ZDR is not available for these “covered models” unless Anthropic expressly authorizes it; otherwise requests from ZDR organizations are rejected with a 400 invalid_request_error (ZDR organizations can enable 30-day retention per workspace). The 30-day retention of prompts and outputs applies wherever covered models are offered — on Bedrock and Google Cloud the retained copy stays with the respective cloud provider.
• New: Enterprise Frontier Safeguards (announced 1 September 2026). Anthropic intends to combine the privacy properties of zero data retention with abuse detection: data stays in the customer’s cloud (AWS, Google Cloud or Azure) under customer-managed keys, and abuse checks run automatically without human review by Anthropic. According to Anthropic the rollout starts in phases “later this fall”; eligible customers are to receive ZDR on Fable 5 and Fable 5.1 until then (on AWS limited until 31 December 2026). Anthropic has not publicly defined which customers qualify — we assess that in the individual project.
• Data residency of the first-party API: the Anthropic API’s inference_geo parameter only accepts the values "us" and "global" — there is no EU option. If processing has to be confined to the EU geographically, the Anthropic API alone cannot deliver that.
• Consequence for EU workloads: this is precisely why we consistently recommend consuming Claude through AWS Bedrock (EU cross-region inference, or in-region Ireland/Stockholm via bedrock-mantle for Opus 4.7+/Sonnet 5/Haiku 4.5) or the Gemini Enterprise Agent Platform (single-region Belgium/europe-west1 or EU multi-region) for GDPR-sensitive processing. There, the data processing rules of the respective cloud provider apply, including the existing data processing agreements — and on Bedrock zero data retention is enabled by default for Opus 5.

Model lifecycle and deprecations

Anthropic now retires older models on a visible cadence. For anyone running production deployments, keeping an eye on the lifecycle belongs in the operating routine:

• Claude Opus 4.1 (claude-opus-4-1-20250805) was retired on 5 August 2026. Anthropic recommends Opus 5 as its successor.
• Claude Fable 5 (claude-fable-5) has been listed as legacy since 1 September 2026; Anthropic recommends migrating to Fable 5.1. Retirement not before 9 June 2027; Fable 5.1 and Mythos 5.1 not before 1 September 2027.
• For the next models Anthropic names earliest possible retirement dates: Sonnet 4.5 not before 29 September 2026, Haiku 4.5 not before 15 October 2026, Opus 4.5 not before 24 November 2026.
• AWS Bedrock and Google Cloud set their own, later retirement dates. If you consume Claude through either hyperscaler, their timelines apply — which in practice buys additional migration headroom.
• Deprecated since Opus 4.7: the sampling parameters temperature, top_p and top_k. Setting them to non-default values makes the API return an error (400). Older integrations that hard-wire these parameters break on a model switch — a common stumbling block in migrations that we regularly find in reviews.

In practice this means a model-ID swap alone is not always enough. We recommend keeping model IDs and request parameters in central configuration instead of scattering them across application code — then switching models stays a configuration change.

Availability in the EU

EU availability of the Claude family is markedly better in 2026 than a year ago. Anthropic shipped Opus 4.7 (April 2026), Opus 4.8 (May 2026), Opus 5 (24 July 2026), Fable 5.1 (1 September 2026) and Opus 5.5 (22 September 2026) on all three third-party platforms on the day of release. Note on the Mythos class: Fable 5.1 and Fable 5 are available in the EU with data residency via the Gemini Enterprise Agent Platform (Europe multi-region), but global only on Bedrock — each with mandatory 30-day data retention; Mythos 5.1 remains restricted to Glasswing participants (see section above).

AWS Bedrock:

• Opus 5.5 (22 September 2026): per the model card only via the EU cross-region inference profile eu.anthropic.claude-opus-5-5 (source regions Frankfurt, Zurich, Stockholm, Milan, Spain, Ireland, London, Paris) or global — no in-region endpoint in Europe; the bedrock-mantle endpoint is in-region only in us-east-1, ap-southeast-4 (Melbourne) and us-gov-west-1. Not a covered model.
• Fable 5.1 (1 September 2026) and Fable 5: available to all Bedrock customers, but according to the model card exclusively via the inference profiles global.anthropic.claude-fable-5-1 and us.anthropic.claude-fable-5-1 — no EU geo profile, no EU in-region endpoint. Frankfurt, Zurich, Stockholm, Milan, Spain, Ireland, London and Paris are listed only as source regions for global cross-region inference. Standard service tier only; data retention runs in aws_review mode (up to 30 days, human review by Amazon personnel).
• Opus 5 (24 July 2026): in-region in eu-west-1 (Ireland) and eu-north-1 (Stockholm) — per the model card exclusively via the bedrock-mantle endpoint (Anthropic Messages API, without Converse/InvokeModel/Guardrails). Via bedrock-runtime, only the EU cross-region inference profile eu.anthropic.claude-opus-5 or global is available there, as from Frankfurt (eu-central-1), Zurich (eu-central-2), Paris (eu-west-3), Madrid (eu-south-2), Milan (eu-south-1) and London (eu-west-2). Routing stays within EU geography either way; zero data retention is enabled by default.
• Opus 4.8: in-region in eu-west-1 (Ireland) and eu-north-1 (Stockholm), likewise only via bedrock-mantle; same EU geo source regions via the profile eu.anthropic.claude-opus-4-8 on bedrock-runtime.
• Sonnet 5 (30 June 2026) and Haiku 4.5: per the model cards, also in-region in eu-west-1 (Ireland) and eu-north-1 (Stockholm) only via bedrock-mantle; via bedrock-runtime only the EU geo profile or global.
• Opus 4.7: in-region in eu-west-1 (Ireland) and eu-north-1 (Stockholm) via bedrock-mantle, otherwise only EU geo/global as above.
• Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5: per the AWS model cards, no in-region hosting in the EU (the bedrock-mantle endpoint does not support these models) — only the EU geo profile (source regions including Frankfurt, Ireland, Stockholm) or global.

Gemini Enterprise Agent Platform (formerly Vertex AI, europe-west1/Belgium and EU multi-region): Per Google’s model pages, Opus 4.5, Opus 4.6, Sonnet 4.5, Sonnet 4.6 and Haiku 4.5 run as single-region in europe-west1 (Belgium) — not Frankfurt/europe-west3. Opus 5.5 (from day one), Fable 5.1, Fable 5, Opus 5, Sonnet 5, Opus 4.8 and Opus 4.7 have no single-region option there and run only via the EU multi-region endpoint (aiplatform.eu.rep.googleapis.com), which exclusively routes within EU geography (10% premium over the global endpoint), or global.

Microsoft Foundry: Claude models exposed as Claude 3P in the model catalogue, GA since early July 2026, including Opus 5 and Sonnet 5 from day one and Opus 5.5 as GA since 22 September 2026 (Global Standard and US Data Zone); Fable 5.1 as a preview since 1 September 2026 (Global Standard only), Mythos 5.1 as a gated research preview with Entra ID authentication. Important for the GDPR assessment: As of 22 September 2026 there is still no EU data zone deployment for Claude in Foundry — only Global Standard and, for the Hosted-on-Azure variants of Opus 5, Opus 4.8 and Sonnet 5 only, Data Zone Standard (US) are offered (hosted in Sweden Central and East US 2, among others), so processing is not restricted to the EU and Anthropic remains the independent data processor. Microsoft lists “Foundry in Europe” as coming in 2026. Until then we recommend Bedrock or the Agent Platform for strict EU residency.

For GDPR-compliant deployments we recommend: Opus 5.5 → Gemini Enterprise Agent Platform EU multi-region (data residency and ML processing in the EU) or the AWS Bedrock EU geo profile from Frankfurt, as the most capable ZDR-capable model; Opus 5 → AWS Bedrock eu-west-1 (Ireland) or eu-north-1 (Stockholm) in-region (only via the bedrock-mantle endpoint/Anthropic Messages API) when a fixed EU location on Bedrock is mandatory; Sonnet 5 → AWS Bedrock or Agent Platform EU multi-region as the cost-efficient default (permanently USD 2 / 10 per 1M tokens since 10 August 2026); Sonnet 4.6 → Gemini Enterprise Agent Platform Belgium (europe-west1) single-region if strict single-region hosting at one fixed EU location is mandatory (on Bedrock, Sonnet 4.6 offers no in-region hosting, only the EU geo profile/global); Opus 4.6 → Gemini Enterprise Agent Platform Belgium (europe-west1) single-region if Opus-tier capability with strict single-region hosting is required (Bedrock likewise has no in-region hosting in the EU for Opus 4.6).

New compliance and enterprise features (July/August 2026)

Beyond the models themselves, Anthropic has shipped a series of platform features since July 2026 that are primarily of interest to compliance, data protection and security owners. Most of them require Claude Enterprise and are in beta:

• Inference hooks (5 August 2026, beta): Every prompt from claude.ai, Cowork and Claude Code can be sent to the company’s own AI security server before inference happens, and that server decides whether to allow or deny the request. This makes it possible to enforce company-specific policies — rules on personal data or trade secrets, for instance — technically, before a prompt ever reaches the model. For organizations that previously could only log after the fact, this is the most relevant addition of the summer.
• Compliance API for Cowork and Claude Code (3 and 11 August 2026, beta): The Compliance API now also returns transcripts of Cowork sessions (in the cloud) and Claude Code sessions (on local machines). That closes a gap for audit, evidence and retention obligations, because agentic work outside the classic chat becomes reviewable too.
• Security scanning for skills and plugins (6 August 2026): Skills and plugins are scanned for security issues for enterprise organizations — relevant wherever staff bring in their own extensions.
• Managed Agents — cost caps and governance (7 August 2026): New additions are session budgets as hard cost ceilings, an additional advisor role, inference_geo pinning per agent, and skills sourced directly from GitHub repositories. An effort setting per agent and webhooks for environment and memory store events had already landed on 22 July 2026.
• Admin API for Claude Enterprise (14 July 2026, beta): User management for Claude Enterprise can be automated — a prerequisite for clean onboarding and offboarding through existing identity processes.
• Claude Cowork on web and mobile (7 July 2026): Cowork is available in the browser and on mobile from the Max plan upwards. The M365 connector gained write access to mail, calendar and files — an extension that deserves a permissions and risk review before rollout, because the agent no longer only reads but also changes data.

Text watermarking from 2 August 2026 (EU AI Act)

For compliance owners in the scope of the EU AI Act, one change from the summer of 2026 is particularly relevant: Anthropic embeds a machine-readable, humanly imperceptible watermark into text generated by Claude. Anthropic cites the EU AI Act and the EU Code of Practice on Transparency of AI-Generated Content, which it signed in July 2026, as the basis.

The key points per Anthropic (post of 14 August 2026):

• Claude models launching in the EU on or after 2 August 2026 support machine-readable marking at launch. For models released earlier the law provides a transition period; Anthropic intends to add watermarking there over the coming months.
• The watermark is woven into the text itself and survives copy-and-paste as well as some editing. Generated files additionally carry digitally signed provenance metadata where supported.
• Per Anthropic there are no extra tokens, so pricing and speed are unaffected, and the quality and content of the output are unchanged.
• Anthropic is rolling the marking out globally because it does not yet have a durable way to scope it by region. A detection API is announced but not yet available.

What this means for your organization: anyone using Claude output in publications, proposals or training material should know that this content remains technically identifiable as AI-generated. That is intentional under Art. 50 of the AI Act and helps rather than hinders deployer obligations — but it belongs in internal AI usage policies so that nobody is caught out by it.

Platform updates from 18 to 20 August 2026

Shortly after our last update, Anthropic moved several features out of beta — relevant above all for teams building on the Claude platform in production:

• Computer use is GA (19 August 2026): the computer use tool leaves beta as the computer_toolset_20260801 toolset — no beta header, batch actions (several actions in one turn), zoom enabled by default, and per-member configuration. Upgrading an existing integration changes the request shape.
• New browser use tool (19 August 2026): browser_toolset_20260801 drives a browser that your own application hosts. It reads the page itself (accessibility tree, elements, forms, tabs) instead of only clicking screenshots, and adds form input, tab management, download reporting and opt-in file upload. Both toolsets are available for Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8.
• Files API, Agent Skills and the Skills API are GA (19 August 2026): the respective beta headers are no longer required. For the Files API, requests without the header use the current response format with file expiration (expires_in_seconds) and pagination.
• Admin API for Claude Enterprise is GA (19 August 2026): members, invites, groups and custom roles can be managed without a beta header — the prerequisite for wiring onboarding and offboarding cleanly into existing identity processes.
• Managed Agents (19 August 2026): web_search and web_fetch can now be restricted to permitted destinations via allowed_domains and blocked_domains — a directly usable governance lever. Sessions running in self-hosted sandboxes can also attach memory stores.
• Workbench is now Playground (18 August 2026): the console surface was renamed and extended; it shows the full SDK request and the API response for each run. Important for data protection documentation: like the Workbench before it, Playground is not covered by zero-data-retention agreements.
• Python SDK v1.0 (20 August 2026): the SDK removes long-deprecated surface — including the legacy Text Completions API and the temperature, top_p and top_k parameters on Messages methods. AnthropicBedrock now also requires an explicitly configured AWS region instead of silently defaulting to us-east-1. If you run Claude through Bedrock in EU Regions, that is welcome news: an unnoticed fallback into a US Region was a realistic failure mode until now.

In practice this means: anyone rolling Claude out more broadly has had markedly better tooling since August 2026 to enforce policies technically and to produce evidence. At the same time the coordination effort grows, because write-capable agents and local coding sessions belong in the data protection and security assessment. We support that assessment as part of our AI compliance consulting.

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. Intelligent model routing helps optimise cost. Simple requests automatically go to Haiku 4.5, complex ones to Sonnet 5 (or Sonnet 4.6 when strict single-region EU hosting is required, e.g. Agent Platform Belgium/europe-west1), and demanding reasoning and agentic workloads (software engineering, long-horizon agents) to Opus 5. This flexibility is exactly what protects against dependence on a single model — if a model is pulled at short notice, as recently happened with Fable 5, CompanyGPT can switch seamlessly to another Claude model or another provider.

Typical deployment scenarios include integration into existing knowledge management systems, connection to ERP systems, or running as an internal chatbot for staff. With CompanyGPT you retain full control over access rights, audit logs and data flows, essential for sectors with high compliance requirements such as banking, insurance, or healthcare.

Our recommendation

For the most demanding GDPR-sensitive workloads: Claude Opus 5 (claude-opus-5). The latest Opus model since 24 July 2026 — per Anthropic close to Fable 5, at half the price and without its 30-day retention mandate. Since 1 September 2026 Claude Fable 5.1 (claude-fable-5-1) offers the highest available capability; because Fable 5.1 as a covered model requires 30-day retention and zero data retention is possible only with Anthropic approval, Opus 5 remains our recommendation for workloads that require ZDR (and on Bedrock ZDR is even the default there). Fable 5.1 pays off for demanding agentic workloads where the 30-day retention is documented as acceptable — in the EU with data residency via the Europe multi-region of the Gemini Enterprise Agent Platform (global only on Bedrock) — per Anthropic the USD 0.25 cache read price makes cache-heavy agents up to 45% cheaper than with Fable 5. Same price as Opus 4.8 (USD 5/25 per 1M tokens), EU hosting in-region on AWS Bedrock in Ireland and Stockholm, from Frankfurt via the EU cross-region inference profile, plus the EU multi-region endpoint of the Gemini Enterprise Agent Platform. Recommended for software engineering, long-horizon agents, multi-agent workflows and high-end reasoning. Existing Opus 4.8 deployments migrate with a model-ID swap at no cost delta.

For most business applications: Claude Sonnet 5 (claude-sonnet-5). Since 30 June 2026 the mid-tier flagship: Anthropic’s strongest Sonnet model, which according to Anthropic significantly closes the gap to the Opus class — and which has cost a permanent USD 2/10 per 1M tokens since 10 August 2026 (the announced increase to USD 3/15 is cancelled). EU-available via AWS Bedrock and now also via the Gemini Enterprise Agent Platform (EU multi-region, 1M token context). Claude Sonnet 4.6 (claude-sonnet-4-6) remains the choice when strict single-region hosting at one fixed EU location is required — available via the Gemini Enterprise Agent Platform in Belgium (europe-west1); on Bedrock, Sonnet 4.6 offers no in-region hosting there (only the EU geo profile/global).

For strict single-region hosting at one fixed EU location with Opus-tier capability: Claude Opus 4.6 (claude-opus-4-6). On AWS Bedrock, no Opus version offers in-region hosting in Frankfurt — only the EU geo profile or global is available there; true in-region hosting on Bedrock only exists from Opus 4.7 onward (incl. Opus 4.8, Opus 5), and then exclusively in Ireland/Stockholm via the bedrock-mantle endpoint (Anthropic Messages API, without Guardrails/Knowledge Base/Agents). Anyone who instead needs strict single-region hosting at one fixed EU location with the full feature set (Converse, Guardrails) will find that for Opus 4.6 via the Gemini Enterprise Agent Platform in Belgium (europe-west1).

For cost-optimised applications: Claude Haiku 4.5 (claude-haiku-4-5). For high volume and simpler tasks (categorisation, simple extraction, content moderation) Haiku 4.5 offers the best price/performance ratio. EU-available across all three third-party platforms.

As innFactory AI Consulting we support you in selecting the right model, the GDPR-compliant integration and cost optimisation for your specific use cases. Contact us for an informal initial consultation.