Models & Providers

DYPAI agents work with DYPAI Managed AI and direct providers such as OpenAI, Anthropic, and Google Gemini. Switching is a single-field change — no code rewrites, no adapter boilerplate. Your tools, memory, and prompts are model-agnostic.

Supported providers

The full list of available models is loaded dynamically in the dashboard when you pick a provider — you'll always see what's currently supported. On DYPAI Managed AI, which models you can pick is plan-gated and managed org-level on the AI Models page (/{org}/ai-models).

DYPAI Managed (no credential)

The fastest path needs no API key at all. Set the provider to DYPAI Managed and pick a model — DYPAI routes the request through its managed (OpenRouter-backed) catalog and bills the usage as AI Credits.

Connect a provider (BYOK)

To bring your own key for OpenAI, Anthropic, or Gemini, save a credential first.

Switching providers

Switching models — even across providers — is just editing the agent node. Nothing else changes:

• Your system prompt stays the same
• Your tools stay the same
• Your memory stays the same (stored as content parts, not provider-specific formats)
• Your frontend code stays the same

This makes A/B testing and fallback strategies trivial. Build once, try any model.

Picking the right model

The model ids below are examples — always check the live dropdown in the dashboard for what's currently available on your plan.

Temperature

The temperature field controls randomness:

• 0.0 – 0.3 → Deterministic, factual. Use for extraction, classification, structured output.
• 0.5 – 0.8 → Balanced. Default for most chat agents.
• 0.9 – 1.2 → Creative, diverse. Use for brainstorming, content generation, varied replies.

Default is 0.7.

Max output tokens

The max_output_tokens field caps how many tokens the model can generate in a single response. Default is 1024. Raise it for long-form answers; lower it to keep responses (and cost) tight. On DYPAI Managed AI this value is clamped by your plan, so very large requests may be capped automatically.

Structured output

Sometimes you want the agent to return JSON matching a specific shape instead of free text — for sentiment analysis, data extraction, tagging, or programmatic workflows.

Pass a JSON Schema in the output_schema field:

{
  "type": "object",
  "properties": {
    "sentiment": {
      "type": "string",
      "enum": ["positive", "negative", "neutral"]
    },
    "confidence": {
      "type": "number",
      "minimum": 0,
      "maximum": 1
    },
    "summary": { "type": "string" }
  },
  "required": ["sentiment", "confidence"]
}

With this set, the agent's response always matches the schema. The response arrives in the object field instead of content:

{
  "object": {
    "sentiment": "positive",
    "confidence": 0.87,
    "summary": "Customer is satisfied with delivery speed."
  }
}

Usage and cost

Every agent response includes a usage block:

{
  "usage": {
    "prompt_tokens": 1820,
    "completion_tokens": 142,
    "total_tokens": 1962
  }
}

How you're charged depends on which mode the agent runs in:

BYOK (your own key). You pay the provider directly — DYPAI does not mark up your tokens. The credential you added is charged by OpenAI / Anthropic / Google per their public pricing.

DYPAI Managed AI. Usage is metered as AI Credits (priced as "AI Credits / 1M tokens"). The agent returns a managed_ai_billing block alongside usage, and the cost is drawn from your monthly plan credits first, then from any purchased top-up packs. There is a credit conversion — it's not pass-through. Free plans get +5 AI Credits/day; Pro gets 100/month. See AI Credits and Managed AI.

Next steps