AI models & suppliers

Overview

Every model below is marked with one of three data-handling classes describing what happens to your prompts and the model's answers at the supplier:

This page is generated from the same registry our application reads, so it reflects the models actually in production. Your contractual rights around these suppliers and models are set out in the Terms & Conditions; how we process personal data is covered by the Privacy Policy and the DPA. Looking for the research connector instead? See the MCP connector documentation.

Data handling classes

Zero Data Retention

The supplier does not store your prompts (Input) or the model's responses (Output) at rest after the response is returned, and does not use them for training. Transient, in-memory processing that exists only for the duration of the request does not count as retention; for our ZDR configuration, prompt caching is additionally disabled and abuse-logging is opted out, and processing is pinned to the EU. Note that ZDR describes the supplier's handling — DirectCase itself still stores your conversation history in the app (deleted after 90 days, per the DPA).

Self-hosted · Zero Data Retention

The model is an open-weights model operated entirely by DirectCase on infrastructure we control, and it is Zero Data Retention as well: your prompts and the model's outputs are processed in memory only for the duration of the request and are never stored at rest, logged or used for training. They never reach a third-party AI model provider. The GPU host provides raw compute only — it has no model-provider role and no access to your data as a service; it is listed as an infrastructure sub-processor in our DPA. As with every class, DirectCase itself still stores your conversation history in the app (deleted after 90 days, per the DPA).

Standard

The model runs on the supplier's API under the supplier's business terms. Suppliers may retain API data for a limited period (typically for abuse monitoring) as described in their linked data-usage policies, but do not use business-tier API data to train their models. If your engagement requires that prompts never be retained by a third-party supplier, choose a Zero Data Retention or Self-hosted model.

Suppliers

The suppliers involved in serving AI features, with links to the terms that govern each of them. Suppliers marked as sub-processors also appear in the sub-processor table of our DPA (Appendix A of the Terms).

Models

The models currently available in the Service, grouped by supplier. The plans a model is available on are shown per model; see Availability by plan for the full matrix.

Internal models

Utility models the Service uses behind the scenes (naming chat threads). They process the same conversation content under the same supplier terms and the same no-training commitments as the chat models.

Where your data lives

Where the models run is only half the picture — this is where DirectCase itself keeps your data:

Application & databases

The DirectCase application, its databases (including your conversation history) and the legal-source corpus run on Hetzner infrastructure in the EU.

Files you upload

Stored in Amazon S3 in the EU.

Backups

Database backups are AES-256-encrypted before leaving our infrastructure and stored in Amazon S3 in the EU.

Self-hosted models

Run on GPU capacity rented from RunPod in the EU. DirectCase operates the model, the inference server and its logs; the GPU host provides compute only.

Prompt caching

Prompt caching stores intermediate computational states so repeated context does not have to be reprocessed. Caching never means your data is shared with other customers — caches are scoped to DirectCase's account with the supplier. The classes used on this page:

Implicit caching

The supplier may transparently cache repeated prompt prefixes to reduce latency and cost. Cache entries are short-lived and scoped to our project.

Caching disabled

All prompt caching is disabled at the project level. Nothing about your request is reused across requests.

In-memory only

The inference server keeps transient key-value states in GPU memory while processing requests. Nothing is written to disk; memory is recycled continuously.

No caching

No prompt caching is used for this model.

Availability by plan

Fallback behaviour

If the standard cloud model you selected is temporarily unavailable, the Service can transparently retry your request with another frontier model so your work is not interrupted. The fallback order is:

1. 1 Gemini 3.8 Flash →
2. 2 GPT-5.6 Terra →
3. 3 Claude 5 Sonnet

• Fallback applies only between the standard cloud models above.
• The Zero Data Retention and self-hosted models never fall back to another vendor — if the model you selected is unavailable, the request fails rather than sending your data outside the boundary you chose.

FAQ

Which model should I choose?

The default (recommended) model is the best balance of quality, speed and cost for legal research. Choose the Zero Data Retention model when a mandate or your bar's guidance requires that prompts are never retained by a third-party supplier; choose the self-hosted model when data must not reach any third-party AI provider at all — it is Zero Data Retention as well, operated by DirectCase. Neither ever falls back to another vendor.

What does Zero Data Retention actually guarantee?

That the supplier stores neither your prompts nor the model's responses at rest after the response is returned, does not use them for training, and — in our configuration — has prompt caching disabled and processing pinned to the EU. It does not change what DirectCase itself stores (your conversation history, deleted after 90 days per the DPA).

Do the AI suppliers train on my data?

No. Our contracts and configurations with every supplier exclude training on your content, and the same prohibition binds DirectCase itself — see the Terms & Conditions and the DPA.

Does the GPU host see my prompts for self-hosted models?

The GPU host rents us hardware; the model weights, the inference server and its logs are operated by DirectCase. The host has no model-provider role. It is listed as an infrastructure sub-processor in our DPA, like our other hosting providers.

What happens when models on this page change?

Adding, removing or upgrading a model from a supplier already listed here is an operational update — this page is updated and nothing changes about your rights. Engaging a new AI supplier that would process your data, or withdrawing Zero Data Retention status from a model, is a material change: we notify affected users by email at least 30 days in advance with a right to terminate, as set out in the Terms & Conditions.