[model] Reference¶
Field reference for the [model] section, the model registry and the session-wide runtime budget. For how to pick and configure models, see the LLM configuration guide.
The registry accepts two mutually exclusive sources. Declare models inline with available_models for simple LiteLLM setups, or point models_python_file at a Python file that exports models: list[BaseLlm] when a model needs custom LiteLLM kwargs such as cache_control_injection_points. Setting both raises an error when the registry loads.
Fields¶
| Field | Type | Description | Default |
|---|---|---|---|
available_models |
list[ModelEntry] |
Models declared inline (see below). Each model value is both the registry key and the id passed to LiteLlm(model=...); duplicates raise. |
[] |
models_python_file |
string |
Path to a Python file exporting models: list[BaseLlm]. Use instead of available_models. |
None |
budget |
float |
Runtime token budget in USD shared by every model call in one session. 0 or omitted means unlimited. |
0.0 |
prices |
list[ModelPrice] |
Custom per-model prices for the runtime budget (see below). | [] |
evaluation_replace_all_models_with_model_name |
string |
When set, evaluation replaces every agent's model with the registered model under this name. None keeps each agent's declared model. |
None |
available_models Entry (ModelEntry)¶
Each entry under [[model.available_models]]:
| Field | Type | Description | Default |
|---|---|---|---|
model |
string |
LiteLLM-style id (<provider>/<model>), also the registry key |
required |
api_key_env |
string |
Environment variable that holds the API key | required |
base_url |
string |
Override the provider base URL | None |
prices Entry (ModelPrice)¶
Each entry under [[model.prices]]. Prices are USD per one million tokens.
| Field | Type | Description | Default |
|---|---|---|---|
model |
string |
Registry key this price applies to | required |
prompt_per_million |
float |
Price per 1M prompt tokens | required |
completion_per_million |
float |
Price per 1M completion tokens | required |
cached_per_million |
float |
Price per 1M cached prompt tokens; defaults to prompt_per_million when omitted |
None |
Example¶
config.toml
[model]
budget = 5.0
[[model.available_models]]
model = "anthropic/claude-sonnet-4-5-20250929"
api_key_env = "ANTHROPIC_API_KEY"
[[model.available_models]]
model = "openai/gpt-5"
api_key_env = "OPENAI_API_KEY"
[[model.prices]]
model = "openai/gpt-5"
prompt_per_million = 1.25
completion_per_million = 10.0