Model providers

Agents run on your own model accounts. An organization connects providers per environment, adds models from them and chooses a default model for each environment.

ProductionDefault model claude-sonnet-4-5

Anthropicanthropic

key checked••••x7Qa

• claude-sonnet-4-5default
• claude-haiku-4-5

AIOpenAIopenai

key checked••••9fK2

• gpt-5
• gpt-5-mini

OpenRouteropenrouter

key checked••••mR41

• openai/gpt-5
• deepseek/deepseek-chat-v3.1

Each environment has its own connections and default. An agent names one of these models, or uses the default.

Connections

A connection is one provider account in one environment: openai, anthropic or openrouter. Its key is checked with the provider before it’s saved, then sealed. Responses only ever show whether it’s set and its last four characters.

A key is checked by listing models, so it must be allowed to: an OpenAI restricted key needs read access to Models.

Models and the default

From a connection you add models from the provider’s list; only models that can call tools are offered. Each environment has a default model. An agent either names one of its environment’s models or uses the default.

Removing a model or connection that the default or an agent uses is refused (409 model_in_use).

What runs see

Every model call names exactly one model, the run’s, and every turn caps its output at 16,000 tokens. When a provider answers with an error, the run reacts to its status:

Provider answers	The run
401	Fails model_key_invalid, and the connection is marked invalid_key.
402, 403	Fails model_failed (out of credits, or refused).
404	Fails model_unavailable, and the model is marked unavailable.
408, 409, 425, 429, 5xx	Retried, honoring Retry-After, then model_unavailable.
A timeout	Retried once, then model_failed.