Section 203 model catalogue
AI models for professionals bound by secrecy
Which models carry German professional secrets when you procure them through us - with the vendor version, the purchase price per million tokens and the date behind every entry.
The route decides, not the model
German law treats a lawyer's, notary's, tax adviser's or doctor's file as a criminal-law secret. Section 203 of the German Criminal Code permits handing it to a participating service provider - an AI provider among them - only if that provider is put under a written confidentiality obligation, and it makes the professional liable if that is missing.
That is why the same model can be usable on one route and unusable on another. This catalogue therefore lists a status per model for procurement through us: whether the chain is closed, whether one piece is still open, or whether the provider excludes something the law requires. The German original with the full route matrix is at ki-modelle-berufsgeheimnistraeger.
Two figures accompany every entry: the vendor version actually running, and the purchase price per million tokens, passed through without a mark-up. Both are maintained in one place and refreshed weekly - this page reads them, it does not keep its own copy.
37 models by status
Grouped by the question a firm actually arrives with: which of these can I use at all?
Available through us
Everything section 203 requires is settled, including the provider's confidentiality undertaking. You do not have to clarify anything further.
- ChatGPT gpt-5.5 EUR 3.16 per million tokens OpenAI · version as of 29 August 2026
- Claude Opus claude-opus-5 EUR 9.20 per million tokens Anthropic · version as of 29 August 2026
- Claude Sonnet claude-sonnet-5 EUR 3.68 per million tokens Anthropic · version as of 29 August 2026
- Claude Haiku claude-haiku-4.5 EUR 1.84 per million tokens Anthropic · version as of 29 August 2026
- Gemini Flash gemini-3.7-flash EUR 0.78 per million tokens Google · version as of 29 August 2026
- Llama llama-4-maverick EUR 0.65 per million tokens Meta · version as of 29 August 2026 Blended from input EUR 0.65 · output EUR 0.65
- Mistral EUR 0.30 per million tokens Mistral AI Blended from input EUR 0.20 · output EUR 0.60
- Mistral Small mistral-small-2603 Mistral AI · version as of 29 August 2026
- Qwen qwen3.8-flash EUR 0.238 per million tokens Alibaba Cloud (Qwen) · version as of 29 August 2026 Blended from input EUR 0.15 · output EUR 0.50
- Gemma gemma-4-26b-a4b-it EUR 0.20 per million tokens Google · version as of 29 August 2026 Blended from input EUR 0.10 · output EUR 0.50
- Apertus EUR 1.44 per million tokens Swiss AI Initiative (EPFL, ETH Zürich) Blended from input EUR 1.25 · output EUR 2.00
- GPT-OSS gpt-oss-120b EUR 0.25 per million tokens OpenAI · version as of 29 August 2026 Blended from input EUR 0.15 · output EUR 0.55
- DeepSeek deepseek-v4-pro-0813 EUR 0.275 per million tokens DeepSeek · version as of 29 August 2026 Blended from input EUR 0.20 · output EUR 0.50 · cache hits EUR 0.10
- Muse Glimmer muse-glimmer-30b Meta · version as of 29 August 2026
- Nemotron nemotron-3.5-lightning NVIDIA · version as of 29 August 2026
- GLM glm-5.3-flash Z.ai (Zhipu AI) · version as of 29 August 2026
- Phi phi-4 Microsoft · version as of 29 August 2026
- ChatGPT mini gpt-5.4-mini OpenAI · version as of 29 August 2026
- ChatGPT kompakt gpt-5.4-nano OpenAI · version as of 29 August 2026
- ChatGPT Reasoning o3 OpenAI · version as of 29 August 2026
- Mistral Large mistral-large-2512 Mistral AI · version as of 29 August 2026
- Mistral Medium mistral-medium-3-5 Mistral AI · version as of 29 August 2026
- Gemini Pro gemini-2.5-pro Google · version as of 29 August 2026
- Gemini Flash Lite gemini-3.5-flash-lite Google · version as of 29 August 2026
Usable, but one piece is still open
You can work with it - but part of the safeguard is open: either the provider does not commit to confidentiality itself, or it depends on an agreement you still have to conclude. Which of the two applies to which route is set out in the full matrix.
- Pharia Aleph Alpha
- Amazon Nova Amazon Web Services
- Amazon Nova kompakt Amazon Web Services
- Mistral kompakt Mistral AI
Not usable for client data
The provider excludes something section 203 requires - usually the confidentiality commitment or the purpose limitation on your inputs. With client data this route does not work.
- Claude Fable claude-fable-5 Anthropic · version as of 29 August 2026
- Claude Mythos Anthropic
Not obtainable through us
In the catalogue because you need to know this as well.
- Microsoft Copilot Microsoft
- NotebookLM Google
- Perplexity Perplexity AI
- Grok grok-4.6 xAI · version as of 29 August 2026
- Le Chat Mistral AI
- DeepL DeepL SE
- Kimi kimi-k3 EUR 1.40 per million tokens Moonshot AI · version as of 29 August 2026 Blended from input EUR 0.86 · output EUR 3.00 · cache hits EUR 0.18
Components, not choices
31 further entries do not answer questions - they make a corpus findable, turn speech into text or read scans. You do not pick them; they work underneath. They are listed because they see the same files, and the same contractual question applies to them.
- Whisper OpenAI
- Voxtral Mistral AI
- Gemini Embedding Google
- Google Text-Embedding Google
- Diktat und Transkription Google
- Gemini Sprachdialog Google
- Mistral OCR Mistral AI
- E5 Embedding Microsoft Research (E5-Modellreihe, offene Gewichte)
- ChatGPT für Code OpenAI
- ChatGPT Modell-Router OpenAI
- OpenAI Embedding OpenAI
- OpenAI Sprachausgabe OpenAI
- OpenAI Bildmodell OpenAI
- Cohere Rerank Cohere
- FLUX Bildmodell Black Forest Labs
- Sprachausgabe Google
- Gemini Bildmodell Google
- Codestral Mistral AI
- Amazon Nova Sprache Amazon Web Services
- Amazon Nova Bildmodell Amazon Web Services
- Amazon Nova Videomodell Amazon Web Services
- Amazon Embedding Amazon Web Services
- Amazon Rerank Amazon Web Services
- Cohere Embedding Cohere
- Qwen für Code Alibaba Cloud (Qwen)
- BGE Embedding BAAI
- Qwen Embedding Alibaba Cloud (Qwen)
- Qwen Rerank Alibaba Cloud (Qwen)
- Mehrsprachiges Embedding Sentence-Transformers (offene Gewichte)
- Jina Embedding Jina AI
- Texterkennung LightOn
Task classes
Firms do not choose a model by name but by task. The German catalogue sorts the models into these 7 classes and names the tasks per profession:
- Document and file summarisation
- Pleading and drafting work
- Contract analysis and clause review
- Research preparation
- Data extraction from receipts
- Correspondence
- Dictation and transcription
The task-by-task breakdown exists in German only for now - see the German catalogue.
Questions we get about this catalogue
What is section 203 of the German Criminal Code?
Why does the route matter more than the model?
What does covered mean?
Does this replace our own professional assessment?
We are not a German firm. Does any of this apply to us?
Which combination fits your work?
We go through your tasks - the files, the client structure, the IT situation - and name the models that carry them, together with the chain they stand on.
Arrange a conversation