Skip to main content
Each Digital Twin picks its own AI model for each job it does. This page explains where you choose them, which model answers when several could, how reasoning effort works, and how to bring your own model.

​
Model uses

A Model Use is one job a model does for your Digital Twin:
  • Conversation (and a fallback conversation model)
  • Image Analysis
  • Image Generation
  • Audio Generation
  • Video Generation
  • Audio Transcription
  • Video Analysis
  • Text-to-Speech
  • Documents Summarization
  • Speech-to-Speech (Convo Mode)
  • Embeddings Generation
  • Moderation
Two more jobs run in the background and are chosen by Praxis AI: KAG analysis (building the knowledge graph from your files) and Praxis Shield (spotting misuse).
A conversation model must support tools and streaming together.

​
Choose models for your Digital Twin

1

Open the Digital Twin

In Admin, open Instances and select Edit on your Digital Twin.
2

Open Personalization and AI Models

The AI Models part of this tab lists one dropdown per Model Use: Conversation & Reasoning, Convo (Live Voice & Avatar Calls), Media and System.
3

Pick a model, or leave Default

— Default — follows the platform default, which Praxis AI keeps current. Pick a model only when you need a specific provider, cost or speed. Save.
The AI Models section of the Personalization and AI Models tab You can also change the conversation model without leaving the chat: open Instance Settings › Conversation (in the chat’s + menu, select Settings, then your Digital Twin’s name). It has the Conversation Model, the Fallback conversation model, and, for admins allowed to add AI models, a way to add a custom model.

​
What the model pickers show

  • The model marked — Default is the platform default for that job.
  • (Deprecated) models are hidden unless you tick Show Deprecated.
  • ⚠ 30-day retention marks a model whose provider keeps data for 30 days (see Zero Data Retention).

​
Per assistant

Each assistant can have its own Conversation Model and STS / Voice Conversation Model. An assistant’s model wins over the Digital Twin’s model while that assistant runs. Choosing a conversation model for an assistant

​
Which model answers

For each job, Pria uses the first of these that is set:
  1. The running assistant’s own model.
  2. A custom AI model of this Digital Twin marked Instance default for this Model Use (see Bring your own AI model).
  3. The model selected for the Digital Twin.
  4. The platform default.
Voice turns can use their own model (for example Anam Response Model, see Realtime Voice & Avatars), and requests through the Chat Completions endpoint can use theirs.

​
Deprecated models

A deprecated model is on its way out. Pria hides it in the pickers and moves any Digital Twin still set to it to a current model — normally the platform default for that job — so the twin keeps working. An assistant that names a deprecated model keeps it until you change it. Review your choices when you see a model marked (Deprecated).

​
Fallback conversation model

A conversation model can fail before it has said anything — the provider is over its usage limit, briefly unavailable, unreachable, or refuses the key. When a Fallback conversation model is set, Pria answers that turn with the fallback instead of showing an error, and the turn’s usage card says which model answered and why the first one could not. Nothing changes when the reply had already started, and a tool that already ran is never run twice.
  • Platform-wide: Praxis AI can set one fallback model for every Digital Twin. None is set unless Praxis AI chooses one.
  • Per Digital Twin: Fallback conversation model, right under the conversation model in the twin editor and in Instance Settings › Conversation. Leave it unset to inherit the platform choice. A twin on its own custom model falls back the same way, and the fallback itself may be a custom model.
  • The fallback runs on its own provider’s credentials, with the same tools translated for that provider. The record of the conversation keeps the model that actually answered; billing follows that model.

​
Platform Models

Every Digital Twin can choose from the same catalog of models from Amazon Bedrock, OpenAI, Anthropic, Google, Mistral, xAI, Kimi, OpenRouter and Stability AI.

Platform Models & Providers

The full catalog — models per provider, status, token limits, capabilities and Thinking support.

​
Reasoning Effort & Thinking

Some models can think before they answer: they spend extra tokens working through the problem. Reasoning Effort sets how much, with five levels that work the same way across providers.

​
How the levels map to each provider

Each provider names thinking differently. Pria translates the level for you; the table shows what each level becomes, so you can predict cost and speed.
Models that don’t think ignore the setting. xAI’s older grok-4.x models and grok-build-0.1 always think, so Pria leaves the parameter out. grok-4.5, grok-4.6 and grok-4.7 think deeply when the parameter is missing, so None sends "low", the closest setting to off; grok-4.3 can switch thinking off, so None means none there, and grok-4.3, grok-4.6 and grok-4.7 take "xhigh" for Max. Gemini 3.7 and 3.8 Flash refuse the near-zero level of earlier Flash models, so None uses their lowest level, "low". Claude Sonnet 5 and Opus 5 think by default, so None sends an explicit “disabled”. Claude Fable 5.1 and Opus 5.5 always think and refuse to switch it off, so None runs them at their lightest setting. Some models have a minimum effort — GPT-6 Astra and GPT-6.1 Sol start at Low, gpt-5.4-pro at Medium — and lower levels are raised to that minimum automatically.

​
Which effort applies

  1. Custom AI Model — a custom model’s own Reasoning Effort wins over everything.
  2. Chat Completion Reasoning Effort — for requests through the Chat Completions Integration.
  3. Reasoning Effort on the Digital Twin.
  4. Platform default: None.
The OpenAI deep research models (o3-deep-research, o4-mini-deep-research) always run at medium effort, whatever the setting.

​
Models that can think

Look for Thinking = “Yes” in the Platform Models tables. Today:
  • Anthropic: Claude Sonnet 5.5, Opus 5.5, Fable 5.1, Sonnet 5, Opus 5, Fable 5, Sonnet 4.6, Haiku 4.5 (Bedrock or Direct API)
  • OpenAI: GPT-6.1 Sol, GPT-6 series (Astra, Sol, Luna), GPT-5.6 series (Sol, Terra, Luna), GPT-5.5, GPT-5.4 series (5.4, 5.4-pro, 5.4-mini, 5.4-nano), GPT-5 series (5.2, 5.1, 5-mini, 5-nano), o4-mini
  • Google Gemini: Gemini 3.8 / 3.6 / 3.5 Flash (Gemini 3.7 Flash was retired by Google in October 2026 — a Digital Twin still set to it answers with Gemini 3.8 Flash), Gemini 3.5 Flash-Lite, Gemini 3.1 Pro Preview, Gemini 3.1 Flash Lite, Gemini 3 Flash Preview, Gemini 2.5 Pro / Flash / Flash Lite
  • Google AI Platform: Gemini 3.8 Flash (the same model on Vertex AI, same thinking levels)
  • xAI: Grok 4.7, Grok 4.6, Grok 4.5, Grok 4.3, Grok 4.20 (reasoning), Grok 4.1 fast (reasoning), Grok Build 0.1
  • Mistral: Mistral Small 4; Magistral Small 2509 on Amazon Bedrock
  • Kimi (Moonshot AI): Kimi K3, Kimi K2.7 code, Kimi K2.6
  • OpenRouter: Claude Opus 5.5, Claude Sonnet 5.5, Claude Sonnet 5, Claude Sonnet 4.6, Gemini 3.8 Flash, Grok 4.7, Grok 4.3 — the same models reached through OpenRouter, with the same thinking levels (“none” runs Claude Opus 5.5, Claude Sonnet 5.5, Gemini 3.8 Flash and Grok 4.7 at their lightest setting and switches Grok 4.3 off)

​
Image & Video Generation Providers

The conversation model calls the image or video tool; Pria sends the request to the model selected for Image Generation or Video Generation.

​
Content Moderation Models

When Enable Moderation is on (see Configuration), every message is checked by the moderation model before the conversation model sees it. Flagged messages are blocked and an email goes to the Digital Twin’s contact address. See Content Moderation for categories, thresholds and false positives.

​
Embeddings Model

Embeddings power all file search: every uploaded file is cut into passages, and each passage is turned into a numeric fingerprint stored in the search index. The Embeddings Generation picker offers one model, text-embedding-3-small (OpenAI, 1,536 dimensions, up to 8,191 tokens per passage).
If the embedding model ever changes, every file already uploaded must be processed again. Two embedding models describe documents in ways that cannot be compared, so old files cannot be found until they are reprocessed.
For chunk size, search depth and the knowledge graph, see Knowledge & RAG Configuration.

​
Prompt Caching

Some providers support prompt caching, which reduces latency and input token costs by reusing previously processed prompt prefixes. Praxis AI enables prompt caching automatically where supported — no configuration is needed.
Prompt caching is most impactful for conversations with long system prompts, many tools, or extended history — exactly the pattern used by Praxis AI’s RAG pipeline. Anthropic and OpenAI caching are enabled by default for all eligible requests.

​
How to Verify Caching is Working

The admin Conversation History view shows per-turn token usage. Look for the cached_tokens (or cache_read_input_tokens for Anthropic) field — a non-zero value confirms the request hit the cache. Cached prefixes save you ~50–90% on those input tokens; you only pay full price for the new portion of each turn.

​
Zero Data Retention (ZDR)

Zero Data Retention (ZDR) means the AI model provider does not store your prompts or the model’s responses after the request completes — nothing you send to the model is retained on the provider’s servers, used for training, or available for later review. ZDR is a contractual and technical guarantee offered by major providers for API traffic, and it is a cornerstone requirement for privacy-sensitive deployments in education, healthcare, and the enterprise. Praxis AI middleware ships only with models that comply with ZDR. Every model in the platform catalog routes through provider API tiers covered by zero-data-retention practices — your organization’s conversations, documents, and knowledge never become provider-side data.

​
The exception: Anthropic Mythos-class models

Anthropic’s Mythos-class models (Claude Mythos 5, Claude Fable 5 and Claude Fable 5.1) are subject to a mandatory 30-day data retention for trust-and-safety purposes, on every platform, effective June 9, 2026. This retention overrides Zero Data Retention agreements — it applies even to organizations with ZDR contracts in place. Because of this, the Claude Fable 5 and Fable 5.1 models carry a ⚠ 30-day retention marker in the model picker. Selecting one is an explicit, informed choice — administrators choosing these models accept that prompts and outputs sent to them are retained by Anthropic for 30 days.
If your organization requires strict ZDR compliance (e.g., for regulatory or contractual reasons), do not select a flagged model. All other catalog models remain fully ZDR-compliant.
For the full provider policy, see Anthropic’s notice: Data retention practices for Mythos-class models.

​
AI-Generated Content Marking

To meet the EU AI Act’s transparency rules (Article 50, in force since 2 August 2026), model providers embed imperceptible, machine-readable marks in what their models generate: Anthropic’s text watermark on Claude (Fable 5.1 now, Opus 5 from 9 September 2026, other Claude models to follow), SynthID on Gemini text, and Content Credentials (C2PA) on generated images from OpenAI, Amazon, Stability AI and Anthropic. The mark is applied inside the model, adds no tokens or characters, changes nothing you can see, and has no effect on credits or latency. Praxis AI relays every output unchanged, so the marks reach your Digital Twin on every channel — there is nothing to configure. See AI-Generated Content Transparency for the per-provider status, how detection works, and what labelling obligations remain with your organization.

​
KAG Analysis Model

When a file is processed for the knowledge graph, a separate KAG analysis model picks out the people, things and relationships in each passage. It is independent of the conversation model, so a heavy conversation model does not raise the cost of building the graph. Praxis AI chooses this model and validates which catalog models fill the graph reliably. It is not a Digital Twin setting: an extractor that fills the graph badly degrades answers quietly rather than visibly. If your files need a different one, contact the Praxis AI team at humans@praxis-ai.com.
Files uploaded before KAG was turned on join the graph only when you process them: under Enable KAG Fusion for your Digital Twin, select Add existing files. See Knowledge & RAG Configuration.

​
Video Analysis Model

When a video is uploaded, Pria transcribes its sound track (with the Audio Transcription model). The Video Analysis model adds a second pass over the picture: it watches the video and describes what was shown on screen — slides, diagrams, demos, on-screen text. That description is placed first, before the transcript, in the processed document, so what was shown becomes searchable and part of the knowledge graph. Only Google Gemini models can watch video. The picker offers:
Video analysis runs at low media resolution for cost control, and is failure-isolated: if the video-analysis pass errors, it is dropped and the transcript is ingested on its own — a video-analysis hiccup never blocks the transcript. Scene analysis is capped at 500 MB / 30 minutes per video; very large videos that exceed the cap or are processed in split segments are transcript-only (no scene description).

​
Chat Completions Endpoint Model

When the Chat Completions Integration is on, outside clients — such as the ElevenLabs Voice Agent — send requests to your Digital Twin through an OpenAI-compatible endpoint. You can give those requests a different conversation model from the one people see in the app. The settings are in the admin editor’s Configuration and Integrations tab, under Chat Completion. Use this when people in the browser should get your best model (say Claude Opus 5.5) but a voice agent calling in should get a faster, cheaper one — in a voice call, waiting is what people notice. Only the conversation model changes; summaries, embeddings, images, audio and speech keep the Digital Twin’s selection.

​
Provider Types

Some model families, such as Claude, are offered through several providers — for example Bedrock, the Anthropic API and OpenRouter. Choose by latency, cost, data residency and which keys you hold.

​
Bring Your Own AI Model (BYOM)

You can connect your own hosted model — for example on Google Vertex AI, behind an OpenAI-compatible gateway, or a custom model on Bedrock — and use it for any Model Use. You need the AI Models add permission.

​
Add a custom model

1

Open AI Models

In Admin, open AI Models. The list shows your Digital Twins’ custom AI models.
2

Add a model

Select Add and fill in the form:
  • Model Name — the exact model id your hosting platform publishes (case sensitive), for example gemini-flash or projects/my-proj/locations/us/models/my-model.
  • Status — Active models are used; Inactive ones are kept but ignored.
  • Description — what the model is, for other admins.
  • Model Use — the job it does (for example Conversation, Image Generation, Documents Summarization).
  • Instance default for this Model Use — tick it to make this model replace the platform model for that job. Untick it to keep the model available only to assistants that pick it (Conversation Model → “Custom models (this instance)”). One default per Model Use.
  • Client Library Type — Open AI (any OpenAI-compatible endpoint, including most gateways), AWS Bedrock, Anthropic, Google GenAI, Mistral AI, X.AI (Grok) or Stability AI.
  • API URL and API Key — your endpoint’s base address and its secret. Keep the key secure and rotate it from time to time.
  • Optional: Max Input Tokens, Max Output Tokens, Reasoning Effort, Extended Context (1M tokens), and the model’s features (Stream, Tools, Vision, MCP…).
3

Save

The model joins the list. A conversation model must support tools and streaming together.
The custom AI model form
Try a custom model on a test Digital Twin first, to check its speed, cost and behaviour before busy twins depend on it.

​
Self-hosted models on RunPod

If your organization runs its own models on RunPod (GPU hosting that starts a model when it is needed), you can connect your RunPod account and use those models as custom AI models.
  1. In AI Models, click Browse RunPod.
  2. Paste your RunPod API key and click Connect. Pria lists the models your RunPod account serves and adds each one as a custom AI model. A model served from several data centres is listed once; Pria sends each request to whichever data centre is ready.
  3. Click Sync now after you add or change models on RunPod. Disconnect forgets the key and makes the synced models inactive; connecting again brings them back.
If you manage several Digital Twins, select one twin before you connect, sync or disconnect — the RunPod account belongs to that twin. Each RunPod model shows whether it is ready: Click Warm up to start a model before people need it. If you stop waiting, Pria withdraws only your own request; a model that someone else is using keeps running. If a model never becomes ready, contact the Praxis AI team at humans@praxis-ai.com.