Choose a model or service
Start with the task and the API your application uses. Then compare available offerings in the live Catalog. An offering's creator alone does not establish its features on Primex.
Start with your task
| You need to | Start here | Check before choosing |
|---|---|---|
| Connect an agent, editor or chat app | Integrations and Setup | The client's API format, required tools and compatible models |
| Generate text or reason through a problem | Chat models | Context/output limits, reasoning controls and response time |
| Produce structured JSON | Structured outputs | Strict schema support for your protocol, rather than JSON text alone |
| Answer questions about documents | Retrieval recipe | Embedding compatibility, reranking and a chat model for the answer |
| Read or extract a document | Document services | Plain OCR, layout or typed fields; formats and model-specific language coverage |
| Create or edit an image | Image models | Generation versus editing, reference inputs, size and output format |
| Transcribe or generate speech | Audio reference | Transcription versus translation, voices, language coverage and input limits |
| Translate, summarize or moderate text | Services | The exact operation, language/default rules and returned result fields |
| Generate a video | Video reference | An enabled offering must exist; the current reference explains the unavailable status |
For document extraction, a general vision model and a specialized invoice/receipt service return different results. For retrieval, an embedding model creates vectors while a reranker orders candidate documents. Choose the operation your application needs before comparing providers.
Check compatibility and limits
- Open the offering's Catalog detail page and copy its exact API ID.
- Confirm the API format and features your workflow needs: streaming, image input, tools, strict JSON or configurable reasoning.
- Check input/output limits and the supported parameters. A large context window does not override current request or token-rate limits.
- Use Setup's compatible selection and first-request example. Add optional controls after the minimal request works.
Keep an embedding collection tied to its model and dimensions. Replacing a model can require re-embedding the collection; compare vectors only in compatible spaces. See the RAG recipe.
Compare on your own workload
Try a small set of representative tasks. Compare answer quality, response time, successful completion and the actual charge in Usage. Include long inputs, files or tool turns when your application uses them.
Prices use different units: input/output tokens, characters, pages, images, megapixels, audio duration or tool calls. A low input-token price alone does not establish the cheapest completed workflow. Reasoning output, cache buckets, long-context pricing and separately metered tools can affect the charge. Use the Catalog's current price and estimator, then inspect the recorded result.
Keep the model ID configurable. Refresh discovery after an unavailable-model error and review withdrawals before moving traffic.
When to read provider docs
Primex Docs explains our endpoints, supported controls, restrictions, responses and billing. Catalog holds current offering facts. Provider docs supply deeper model and feature guidance:
| Reference | Useful for |
|---|---|
| OpenAI model documentation | Model families, prompting and native feature background |
| Claude model documentation | Claude model behavior and provider-native features |
| Azure Document Intelligence | Document models, extracted field definitions and recognition coverage |
| Mistral models | Model and document-understanding guidance |
| Cohere models | Embedding and reranking model background |
| Black Forest Labs FLUX guidance | Image-model behavior and prompting |
For detailed language lists, follow the offering's Catalog language-support links or the relevant service category. Those links are specific to the operation where possible.
Use Primex's documented contract when making Primex requests. A feature in a provider's direct API may be absent on its hosting route or on Primex. For example, our document translation accepts DOCX up to 2 MiB, even though Azure describes more formats. Our Files API IDs have their own ownership and lifetime. Continue using your Primex key and endpoint; provider credentials are not required to call a documented Primex operation.