feat(anthropic): dynamic mapping of Max Tokens for Anthropic (#31946)

- **Description:** Dynamic mapping of `max_tokens` as per the choosen
anthropic model.
- **Issue:** Fixes #31605

@ccurme

---------

Co-authored-by: Caspar Broekhuizen <[email protected]>
Co-authored-by: Mason Daugherty <[email protected]>

Author: 3 people

libs/partners/anthropic/langchain_anthropic/chat_models.py

@@ -7,7 +7,7 @@
 from collections.abc import AsyncIterator, Iterator, Mapping, Sequence
 from functools import cached_property
 from operator import itemgetter
-from typing import Any, Callable, Literal, Optional, Union, cast
+from typing import Any, Callable, Final, Literal, Optional, Union, cast
 
 import anthropic
 from langchain_core._api import beta, deprecated
@@ -61,6 +61,32 @@
 }
 
 
+_MODEL_DEFAULT_MAX_OUTPUT_TOKENS: Final[dict[str, int]] = {
+    "claude-opus-4-1": 32000,
+    "claude-opus-4": 32000,
+    "claude-sonnet-4": 64000,
+    "claude-3-7-sonnet": 64000,
+    "claude-3-5-sonnet": 8192,
+    "claude-3-5-haiku": 8192,
+    "claude-3-haiku": 4096,
+}
+_FALLBACK_MAX_OUTPUT_TOKENS: Final[int] = 4096
+
+
+def _default_max_tokens_for(model: str | None) -> int:
+    """Return the default max output tokens for an Anthropic model (with fallback).
+
+    Can find the Max Tokens limits here: https://docs.anthropic.com/en/docs/about-claude/models/overview#model-comparison-table
+    """
+    if not model:
+        return _FALLBACK_MAX_OUTPUT_TOKENS
+
+    parts = model.split("-")
+    family = "-".join(parts[:-1]) if len(parts) > 1 else model
+
+    return _MODEL_DEFAULT_MAX_OUTPUT_TOKENS.get(family, _FALLBACK_MAX_OUTPUT_TOKENS)
+
+
 class AnthropicTool(TypedDict):
     """Anthropic tool definition."""
 
@@ -1205,7 +1231,7 @@ def get_weather(location: str) -> str:
     model: str = Field(alias="model_name")
     """Model name to use."""
 
-    max_tokens: int = Field(default=1024, alias="max_tokens_to_sample")
+    max_tokens: Optional[int] = Field(default=None, alias="max_tokens_to_sample")
     """Denotes the number of tokens to predict per generation."""
 
     temperature: Optional[float] = None
@@ -1343,6 +1369,15 @@ def _get_ls_params(
             ls_params["ls_stop"] = ls_stop
         return ls_params
 
+    @model_validator(mode="before")
+    @classmethod
+    def set_default_max_tokens(cls, values: dict[str, Any]) -> Any:
+        """Set default max_tokens."""
+        if values.get("max_tokens") is None:
+            model = values.get("model") or values.get("model_name")
+            values["max_tokens"] = _default_max_tokens_for(model)
+        return values
+
     @model_validator(mode="before")
     @classmethod
     def build_extra(cls, values: dict) -> Any:

libs/partners/anthropic/tests/cassettes/TestAnthropicStandard.test_stream_time.yaml.gz

@@ -901,7 +901,10 @@ class color_picker(BaseModel):
 
 @pytest.mark.vcr
 def test_web_search() -> None:
-    llm = ChatAnthropic(model="claude-3-5-sonnet-latest")  # type: ignore[call-arg]
+    llm = ChatAnthropic(
+        model="claude-3-5-sonnet-latest",  # type: ignore[call-arg]
+        max_tokens=1024,
+    )
 
     tool = {"type": "web_search_20250305", "name": "web_search", "max_uses": 1}
     llm_with_tools = llm.bind_tools([tool])

libs/partners/anthropic/tests/integration_tests/test_chat_models.py

@@ -111,6 +111,45 @@ def test_anthropic_proxy_from_environment() -> None:
         assert llm.anthropic_proxy == explicit_proxy
 
 
+def test_set_default_max_tokens() -> None:
+    """Test the set_default_max_tokens function."""
+    # Test claude-opus-4 models
+    llm = ChatAnthropic(model="claude-opus-4-20250514", anthropic_api_key="test")
+    assert llm.max_tokens == 32000
+
+    # Test claude-sonnet-4 models
+    llm = ChatAnthropic(model="claude-sonnet-4-latest", anthropic_api_key="test")
+    assert llm.max_tokens == 64000
+
+    # Test claude-3-7-sonnet models
+    llm = ChatAnthropic(model="claude-3-7-sonnet-latest", anthropic_api_key="test")
+    assert llm.max_tokens == 64000
+
+    # Test claude-3-5-sonnet models
+    llm = ChatAnthropic(model="claude-3-5-sonnet-latest", anthropic_api_key="test")
+    assert llm.max_tokens == 8192
+
+    # Test claude-3-5-haiku models
+    llm = ChatAnthropic(model="claude-3-5-haiku-latest", anthropic_api_key="test")
+    assert llm.max_tokens == 8192
+
+    # Test claude-3-haiku models (should default to 4096)
+    llm = ChatAnthropic(model="claude-3-haiku-latest", anthropic_api_key="test")
+    assert llm.max_tokens == 4096
+
+    # Test that existing max_tokens values are preserved
+    llm = ChatAnthropic(
+        model="claude-3-5-sonnet-latest", max_tokens=2048, anthropic_api_key="test"
+    )
+    assert llm.max_tokens == 2048
+
+    # Test that explicitly set max_tokens values are preserved
+    llm = ChatAnthropic(
+        model="claude-3-5-sonnet-latest", max_tokens=4096, anthropic_api_key="test"
+    )
+    assert llm.max_tokens == 4096
+
+
 @pytest.mark.requires("anthropic")
 def test_anthropic_model_name_param() -> None:
     llm = ChatAnthropic(model_name="foo")  # type: ignore[call-arg, call-arg]