Standard Model Exceptions — retry & handle any provider

What it does

LangChain 1.3.16 introduces standard model exception types in langchain-core (PR #39538). Instead of catching a generic Exception or branching on each vendor's error classes, you now get a small, consistent hierarchy that every model integration — OpenAI, Anthropic, Fireworks, Perplexity, and the rest — raises for the same class of failure: ModelError (base), ModelRateLimitError, ModelTimeoutError, ModelAuthenticationError, and ModelNotFoundError. Because it ships in langchain-core, every partner package inherits it automatically — which is why it shows up in the changelogs of langchain, langchain-openai, and langchain-fireworks on the same day.

Why it matters

In production you almost always wrap model calls in retry and error-handling logic. Before this, you had to branch on each vendor's error classes (openai.RateLimitError, anthropic.APIStatusError, ...) or catch a broad Exception and guess. Standard exception types give you one portable except clause that works no matter which model provider you swap in. That is invaluable in multi-provider apps (fallback chains, model routing) or a library other teams consume — your error handling stops leaking provider-specific types. Paired with the ModelRetryMiddleware fix in the same release (re-raising non-retryable exceptions), you get clean, provider-agnostic resilience logic.

Example

$ safe_generate — translate any failure into a standard exception

from langchain_core.exceptions import (
    ModelError,
    ModelRateLimitError,
    ModelTimeoutError,
    ModelAuthenticationError,
    ModelNotFoundError,
)
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

def safe_generate(prompt: str) -> str:
    try:
        return llm.invoke(prompt).content
    except ModelRateLimitError as e:
        print(f"[retryable] rate limited: {e}")
        raise
    except ModelTimeoutError as e:
        print(f"[retryable] timed out: {e}")
        raise
    except ModelAuthenticationError as e:
        print(f"[fatal] bad API key: {e}")
        raise
    except ModelNotFoundError as e:
        print(f"[fatal] model does not exist: {e}")
        raise
    except ModelError as e:
        print(f"[unknown] {e}")
        raise

# Works identically no matter which provider backs llm
try:
    print(safe_generate("Say hello in one word."))
except ModelError:
    print("Handled a standard model exception.")

Common flags

ModelError

Base class for all standard model exceptions

ModelRateLimitError

Provider throttles requests — retryable

ModelTimeoutError

Call exceeds the timeout — retryable

ModelAuthenticationError

Invalid/missing credentials — not retryable

ModelNotFoundError

The requested model does not exist — not retryable