CompanyGPT vs. Claude: Anthropic directly or through your own cloud tenant?

This page answers a question we are asked regularly in projects: we already use Claude and we are happy with it — do we even need a platform on top of it?

The answer depends less on the model than on the route by which you obtain it. Claude is the same model in both cases. What differs is who processes the data, where that happens and how it is billed.

The two routes in brief

Claude directly from Anthropic (Team, Enterprise)

You sign the contract with Anthropic and use Claude through claude.ai, the desktop apps or the Anthropic API. The interface is mature, the models are available on the day they are released, and setup takes minutes. Anthropic is your processor in this case.

Claude 3P through your own cloud tenant

“3P” stands for third-party provider: you obtain the same Claude models through AWS Bedrock, the Google Gemini Enterprise Agent Platform (formerly Vertex AI) or Microsoft Foundry. The contractual partner for the processing is then the cloud provider with whom you already have a DPA in place. CompanyGPT sits on top of that as the platform layer and brings the interface, roles, knowledge base and cost transparency with it.

The decisive difference: where is the data processed?

A data processing agreement governs who processes data and under what conditions. It does not automatically determine where that happens. This is exactly where the two routes diverge.

With the direct route, Anthropic is your processor. Which regions are contractually assured for the place of processing is settled in the enterprise contract — via the third-party providers the region is part of your existing cloud DPA. For many use cases either is unproblematic. But as soon as professional secrets, HR data or client data are involved, the place of processing becomes a question you have to document — and then a route with a contractually assured EU region makes for a much easier conversation with your data protection officer.

Across the third-party providers the picture is more nuanced, and the differences are bigger than many expect:

So if you need strict EU residency, Bedrock or the Google platform are currently the ones to choose. On Foundry it is, according to the region availability documented there (as of 22 August 2026), not yet covered for Claude, even though the models are listed in its catalogue.

A detail that is often overlooked: data retention per model

The retention rules differ within the Claude family, and for regulated industries that is the most important point on this page:

According to Anthropic’s data-retention information (as of August 2026):

• Claude Opus 5 is available as a zero data retention model; on Bedrock, ZDR is enabled by default.
• Claude Fable 5 is subject to a 30-day retention period and is not listed there under zero data retention — not even for EU endpoints.

For workflows with special confidentiality obligations — law firms, clinics, tax advisors, § 203 StGB contexts — we therefore recommend Opus 5 and use Fable 5 only where the 30-day retention is demonstrably justifiable on the record. In an interface that simply offers “Claude”, this distinction is quite simply lost.

What the route means for cost

An honest word belongs here, because the maths does not come out cheaper in every direction.

A subscription is billed per user per month. Third-party access is billed per token, at the list price of the respective hyperscaler. In our experience from customer projects, token costs under intensive use of the strong models are noticeably higher than what a subscription costs per head — so Claude 3P is not automatically the cheaper route.

That is precisely why we consider the combination sensible rather than the either/or decision: CompanyGPT as the foundation with multi-model routing, Claude 3P for the cases that genuinely need Claude. Most day-to-day work — summaries, translations, standard answers from the knowledge base — runs more cheaply on smaller or open models. The demanding tasks you route to Claude. To keep that controllable, companyDASHBOARD shows consumption and cost per user, team and model.

Incidentally, with us tokens are billed at the hyperscaler’s list price, with no surcharge by innFactory. You pay for the platform, not for the models.

What the direct route does not give you

Anthropic delivers an excellent model and a good interface for individual work. What enterprises need beyond that is not part of the model — and that layer is exactly what CompanyGPT brings:

• A central knowledge base across your documents with companyRAG, including mirroring of SharePoint permissions
• Roles and access rights along your organisational structure instead of per individual account
• The Office add-in from companyM365: CompanyGPT as a task pane directly in Word, Excel, PowerPoint and Outlook — without a Copilot licence per user
• Model variety: Claude, GPT, Gemini and open models in parallel, routed by use case
• Workflow automation via n8n and the open Model Context Protocol
• Cost and usage transparency via companyDASHBOARD
• Operation in your own cloud tenant, on Azure, Google Cloud or sovereign on STACKIT

When the direct route is the right choice

We say openly when you do not need a platform:

• Small teams without particular compliance requirements. If five of you work on non-critical content, a subscription gets you there faster and more simply.
• Pure developer teams that work through the API anyway and need no interface for business departments.
• When it is exclusively about Claude and model variety plays no role.

But as soon as several departments are involved, company knowledge is to be integrated, the place of processing has to be documented or several models make sense, the maths tips in favour of the platform.

Conclusion

This is not Claude versus CompanyGPT — CompanyGPT runs Claude. The question is whether you obtain the model directly from the maker or through your own cloud tenant, and which layer you need on top of it.

For companies with data subject to documentation requirements, the route through a third-party provider in an EU region is the more robust one. For companies with several departments and mixed use cases, the platform layer comes on top. And because tokens are not cheap, the economically sensible option is usually the combination: everyday load on cheaper models, Claude where it makes the difference.

Sources

Retrieved on 23 August 2026:

• Anthropic: zero data retention and model availability — https://support.claude.com/en/articles/15425996
• Claude on Amazon Bedrock — https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html
• Gemini Enterprise Agent Platform — locations — https://docs.cloud.google.com/gemini/enterprise/docs/locations
• Microsoft Foundry — region availability — https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability

Note on the information given: All statements about other vendors’ products are based on their publicly available documentation as of the date stated. Vendors continuously develop their products, features and terms — the vendor’s current information is always authoritative. If any statement appears outdated or inaccurate to you, drop us a line at info@innfactory.ai; we will check and correct it promptly. This comparison is not a substitute for legal or data protection advice in an individual case.

Further reading

• Using Claude in the enterprise – the routes in detail, drawn from more than 50 real-world projects
• Claude in the model overview – versions, benchmarks and EU availability
• GDPR-compliant AI platforms compared 2026 – the market overview
• CompanyGPT vs. Microsoft Copilot – when Microsoft 365 is the starting point