Choose an API format and model
Use the format implemented by your SDK or tool and an offering that publishes support for it. The same Primex key authenticates these paths, but their request and response shapes differ.
Base URL versus request path
| Client | Set this base URL | Client appends | Final request |
|---|---|---|---|
| OpenAI SDK: chat | https://gateway.primexos.com/v1 | /chat/completions | /v1/chat/completions |
| OpenAI SDK: Responses | https://gateway.primexos.com/v1 | /responses | /v1/responses |
| Anthropic SDK / Claude Code | https://gateway.primexos.com | /v1/messages | /v1/messages |
| Direct HTTP | Use the complete URL in the reference | Nothing | The documented operation path |
Clients using another provider adapter may append a different suffix. Follow
that integration's guide; do not add a second /v1. Native image, audio,
document, language and safety operations use their own request contracts.
Discover IDs or inspect capabilities
SDK/editor discovery: GET /v1/models requires your API key and returns
OpenAI-compatible identity fields. It defaults to available chat, code and vision
models. A service_type filter selects another type. This non-billable read is
available at zero balance; it does not return pricing or a capability matrix.
Application catalogue: the public
https://app.primexos.com/api/public/catalog response's offerings list includes
public IDs, types, availability, supported protocols, billing metadata and customer
pricing. Use this to build rich pickers. Setup
filters current offerings for its documented integration contracts.
See Model Discovery and Models.
Check the features your request needs
Native Responses preserves Responses reply/event formats. The legacy rewrite for a chat-only model uses Chat Completions shapes and does not qualify a Responses SDK or agent. Messages has native and reviewed translated paths; translated support does not imply native thinking, Files or Skills. Use Responses and Messages for those boundaries.
Check the selected offering's published protocol and feature support, not its creator name. In particular:
- Chat Completions support does not imply native Responses support.
- Messages support in an SDK does not certify every coding-agent workflow.
- Image input, tools, structured output, reasoning controls and sampling parameters can have model-specific restrictions.
- Native service language/default rules are operation-specific. OCR's recognition locale is different from translation's target language.
- Gated provider-run tools are separate from ordinary client-executed functions.
Start with the minimal documented request and omit optional sampling settings. Add supported options as needed. On an unavailable-model response, refresh discovery and update saved selections rather than retaining a hard-coded inventory.
See Parameters and compatibility for control locations,
output limits and X-Primex-Warnings. Use
Conversation history and state for the history and
tool-result items your application must retain.
Bring an existing application
Follow the OpenAI migration or OpenRouter migration guide. Verify streaming, tools and output handling in addition to changing the key and base URL.