Models and providers

Promptic lets you choose models from supported providers for experiments, evaluations, and tool workflows. Configure access centrally, then select from the models available to your Organization or AI Application.

Managed model access

Managed access uses provider credentials operated by Promptic. It is the fastest setup and creates separate managed-model usage in addition to Promptic Credits. Available models can vary by plan, region, provider availability, and organization policy.

Use managed access when you want a supported default and do not need the calls billed directly to your provider account.

Managed-model processing regions vary. The model picker shows the processing region for each model, so verify the displayed region before selecting a model for workloads with residency requirements.

Bring your own key

Organization administrators can configure provider credentials under Organization settings → Models. Current first-class options include:

• OpenAI API key;
• OpenRouter API key and model enablement;
• Google AI Studio API key;
• Google Vertex AI service account, project, and location.

When you select a BYOK model, all target-model predictions run through your configured provider credential. Promptic's optimizer-model calls still use managed models. An OpenAI API key is the exception: it can also be used for supported optimizer calls, enabling a fully BYOK prompt optimization workflow.

BYOK model cost is billed by the provider. Platform work can still consume Promptic Credits.

Keys are write-only after saving. Replace a key instead of expecting the dashboard to reveal it.

Custom providers

Add a custom provider when it exposes an OpenAI-compatible endpoint. Provide a display name, base URL, API key, and explicit model names. Promptic tests the connection before saving it.

Custom providers can be configured for an Organization or as a narrower AI Application override. Use the narrowest scope that matches who should use the credential.

All target-model predictions for a selected custom model run through that custom provider. Promptic's optimizer-model calls continue to use managed models; custom providers do not currently provide a fully BYOK optimization workflow.

Compatibility is not the same as quality. A successful connection test proves that the endpoint accepts the expected API shape; it does not prove that every structured-output or tool-call workflow behaves identically to a first-class provider.

Model availability

Organization administrators can disable individual models without deleting provider credentials. Use this to enforce internal policy or reduce accidental use of expensive models.

When an AI Application has its own model settings, they override the organization default for that application. Record intentional overrides so an experiment can be reproduced later.

Region and data processing

Do not infer data residency from a model name alone. The actual processing location depends on the endpoint, provider account configuration, routing product, and contractual terms.

Before using a provider for restricted data, verify:

1. the exact endpoint Promptic will call;
2. the provider’s processing and retention guarantees for your account;
3. whether routing or fallback can leave the intended region;
4. whether managed access or BYOK is being used;
5. whether prompts, outputs, and traces should be captured at all.

See Security and data handling for trace payload guidance.

Choosing defaults

• Start with a small model for setup and test experiments.
• Validate model changes against representative examples and the same evaluation criteria.
• Keep a held-out evaluation split to detect overfitting.
• Compare provider changes on the same dataset cases and evaluators.

Model calls can incur cost. Review Billing and pricing before starting large grids or LLM-judge evaluations.