Fallback Models

from agno.agent import Agent
from agno.models.anthropic import Claude
from agno.models.openai import OpenAIChat

agent = Agent(
    model=OpenAIChat(id="gpt-5.4-mini"),
    fallback_models=[Claude(id="claude-sonnet-4-5-20250929")],
)

fallback_models is shorthand for FallbackConfig(on_error=...). When the primary model raises an eligible ModelProviderError, Agno tries the fallback models in order until one succeeds. Install both provider integrations and set OPENAI_API_KEY and ANTHROPIC_API_KEY before running the examples:

uv pip install "agno[openai,anthropic]"
export OPENAI_API_KEY="your-openai-api-key"
export ANTHROPIC_API_KEY="your-anthropic-api-key"

Model strings work for the primary and fallback models:

from agno.agent import Agent

agent = Agent(
    model="openai-chat:gpt-5.4-mini",
    fallback_models=["anthropic:claude-sonnet-4-5-20250929"],
)

Fallbacks catch model-provider errors. Exceptions outside ModelProviderError, such as application runtime errors, propagate without trying another model.

Fallback configuration wraps calls to the agent or team leader’s primary model. It does not wrap separate reasoning_model, parser_model, or output_model calls.

Route Errors to Different Models

from agno.agent import Agent
from agno.models.anthropic import Claude
from agno.models.fallback import FallbackConfig
from agno.models.openai import OpenAIChat

agent = Agent(
    model=OpenAIChat(id="gpt-5.4-mini"),
    fallback_config=FallbackConfig(
        on_rate_limit=[
            OpenAIChat(id="gpt-4.1-mini"),
            Claude(id="claude-sonnet-4-5-20250929"),
        ],
        on_context_overflow=[
            Claude(id="claude-sonnet-4-5-20250929"),
        ],
        on_error=[
            Claude(id="claude-sonnet-4-5-20250929"),
        ],
    ),
)

Agno classifies the primary error and selects one fallback list:

Provider failure	Selected list	Behavior when the list is empty
Rate limit or overload, including status 429 or 529	`on_rate_limit`	Use `on_error`
Context-window overflow	`on_context_overflow`	Use `on_error`
Other server, connection, or provider error	`on_error`	Re-raise the primary error
Other HTTP 4xx client error	None	Re-raise the primary error

Context-window errors are detected from ContextWindowExceededError or recognized error-message patterns. Other 4xx errors, including 400, 401, 403, 404, 413, and 422, are not routed through on_error. If every model in the selected list raises ModelProviderError, Agno re-raises the original primary-model error. The list is selected once from the primary error. A later fallback failure does not select a different error-specific list.

Retries and Fallback

Each model applies its own retry settings before its failure reaches the fallback layer:

from agno.agent import Agent
from agno.models.anthropic import Claude
from agno.models.openai import OpenAIChat

agent = Agent(
    model=OpenAIChat(
        id="gpt-5.4-mini",
        retries=3,
        exponential_backoff=True,
    ),
    fallback_models=[
        Claude(id="claude-sonnet-4-5-20250929", retries=2),
    ],
)

retries=3 allows up to four primary attempts: the first attempt and three retries. The fallback allows up to three attempts with retries=2. Retries apply only to retryable errors. Context-window errors and non-retryable client errors skip the remaining retry loop. An eligible context-window error can therefore activate a fallback immediately.

Streaming

Fallbacks also wrap streaming model calls. When the primary model activates a fallback, Agno resets the accumulated response content before processing the first fallback’s chunks.

A mid-stream failure can occur after primary-model or fallback-model chunks have reached the consumer. Those emitted chunks cannot be retracted. If one fallback emits chunks and then fails, Agno does not reset accumulated content before trying the next fallback. The completed run can therefore combine partial output from the failed fallback with output from the successful fallback. Treat streamed output as provisional, or use a non-streaming run when failover must produce one atomic response.

Fallback Callback

Use callback to record which fallback completed the request:

from agno.agent import Agent
from agno.models.anthropic import Claude
from agno.models.fallback import FallbackConfig
from agno.models.openai import OpenAIChat


def on_fallback(primary_model_id: str, fallback_model_id: str, error: Exception) -> None:
    print(f"[fallback] {primary_model_id} -> {fallback_model_id} (reason: {error})")


agent = Agent(
    model=OpenAIChat(id="gpt-5.4-mini"),
    fallback_config=FallbackConfig(
        on_error=[Claude(id="claude-sonnet-4-5-20250929")],
        callback=on_fallback,
    ),
)

The callback runs after a fallback succeeds. For streaming calls, it runs after the fallback stream completes. Callback exceptions are ignored so they do not interrupt a successful fallback.

Teams

Fallback configuration on a Team applies to the team leader’s model calls. Configure member agents separately when they also need fallbacks.

from agno.agent import Agent
from agno.models.anthropic import Claude
from agno.models.openai import OpenAIChat
from agno.team import Team

researcher = Agent(
    name="Researcher",
    role="Research the requested topic.",
    model=OpenAIChat(id="gpt-5.4-mini"),
)

team = Team(
    name="Research Team",
    model=OpenAIChat(id="gpt-5.4-mini"),
    fallback_models=[Claude(id="claude-sonnet-4-5-20250929")],
    members=[researcher],
)

Parameters

Available on Agent and Team:

Parameter	Type	Description
`fallback_models`	`List[Model \| str]`	Models used as the `on_error` fallback list.
`fallback_config`	`FallbackConfig`	Error-specific fallback lists and callback. Takes precedence when both parameters are set.

FallbackConfig fields:

Field	Type	Default	Description
`on_error`	`List[Model \| str]`	`[]`	General fallback list for eligible provider errors.
`on_rate_limit`	`List[Model \| str]`	`[]`	Preferred list for rate-limit and overload errors.
`on_context_overflow`	`List[Model \| str]`	`[]`	Preferred list for context-window errors.
`callback`	`Callable[[str, str, Exception], None]`	`None`	Called after a fallback succeeds with the primary ID, fallback ID, and primary error.

Get Started

Basics

Advanced

Production

Providers

Other

Additional Resources

Fallback Models

Route Errors to Different Models

Retries and Fallback

Streaming

Fallback Callback

Teams

Parameters

Developer Resources

​Route Errors to Different Models

​Retries and Fallback

​Streaming

​Fallback Callback

​Teams

​Parameters

​Developer Resources

Route Errors to Different Models

Retries and Fallback

Streaming

Fallback Callback

Teams

Parameters

Developer Resources