Provider matrix

This matrix describes built-in routing and wire adapters in the current source. It does not promise account entitlement, funding, or availability of every model documented by a provider.

Id	Adapter	Built-in API family	Credential	Catalog note
glm	OpenAI-compatible Chat	GLM standard Chat Completions	GLM_API_KEY	Versioned static official schema manifest; no guessed list-models call
anthropic	Anthropic Messages	Anthropic v1	ANTHROPIC_API_KEY	Availability depends on credential and endpoint behavior
openai	OpenAI Responses	OpenAI v1	OPENAI_API_KEY	Uses the Responses adapter
deepseek	OpenAI-compatible Chat	DeepSeek v1	DEEPSEEK_API_KEY	Provider-specific error profile
minimax	OpenAI-compatible Chat	MiniMax compatible root	MINIMAX_API_KEY	Provider-specific error profile
fireworks	OpenAI-compatible Chat	Fireworks inference v1	FIREWORKS_API_KEY	Provider-specific error profile

Default selection

Fresh sessions first reuse a validated last-success route when one exists and no operator route was supplied. Otherwise they prefer the built-in openai route when it has a locally usable credential; if not, Iteron selects the first locally credentialed provider. An explicit CLI, environment, or trusted user route is not replaced. No provider or model is assumed to be available to every account. For the built-in OpenAI Responses route, Iteron prefers Codex v0.156.0's API-visible model order only among models admitted by the live catalog; gpt-6-astra is first when available. Other providers retain their documented or catalog selection behavior. Explicit model choices remain authoritative.

Availability states

The selection UI distinguishes usable, unavailable, disabled, and unknown evidence. In particular:

• schema or catalog presence is not account entitlement;
• a private deployment without a healthy default is not guessed;
• a compatible gateway without a model-list endpoint needs an operator manifest or explicit model;
• provider errors can update account or model health;
• a grey entry is intentionally not selectable.

Capability evidence

Reasoning effort, context limits, caching, and other model capabilities are attached only where the exact endpoint/model evidence supports them. A family neighbor does not inherit the limits documented for another model.

The catalog implementation is evolving. Use /model and /status as the runtime view, and report stale or contradictory provider evidence with a synthetic reproduction.