Which model should I name?
Model ids carry their provider, the set of callable models is a live fact rather than a documented list, and models arrive and retire on their providers’ schedules rather than ours.Models
Id forms, how to discover what your key can call, and how a rate is expressed.
Model capabilities
Which features a given model supports, and how to ask before you rely on one.
Model lifecycle
How a model appears, changes, and goes away, and what your code sees when it does.
What can I put in a request?
These are the request-shaping features. Each is the OpenAI parameter you already know, with a note on what differs when a router rather than a single provider is serving it.Streaming
Server-sent events, the usage comment before the final line, and mid-stream failure.
Tool calling
Passing tools through, and what a model that cannot call them does instead.
Structured outputs
JSON mode and schema-constrained responses, and which models honour them.
Image inputs
Multi-modal messages, and the refusal a text-only model returns.
Reasoning
Reasoning-model behaviour, the tokens it spends, and how they are billed.
Prompt caching
Where caching happens, who controls it, and what it does to a request’s cost.
Which endpoint do I call?
Chat completions is the one most people start with. Four more serve on the same key, the same model ids and the same balance.Responses API
OpenAI’s Responses format, and the stored-state fields we refuse because we keep nothing.
Completions and embeddings
The legacy text completion endpoint, and where embeddings stand: not served yet.
Anthropic Messages API
The Messages wire for Anthropic SDKs and harnesses, with either credential header.
Usage lookup
Read one request’s tokens and charge back with the key that made it.
What happens when something goes wrong?
Routing is the part of the product you did not have when you called a provider directly, so it is the part worth reading before an incident rather than during one. Guardrails are the part that stops a key before it becomes an incident.Routing profiles
How a request reaches a provider, and what you can say about that choice.
Bring your own key
Serving a model on your own provider credential instead of ours.
Guardrails
A spend cap, a data-retention rule and prompt checks, attached per key.
Errors
The error envelope, the closed set of types, and what each one asks you to do.