feat: add support for opus 4.5 (#6256)

* feat: add support for new model

* fix: just stage-api && just publish-api (anthropic model settings changed)

* fix: just stage-api && just publish-api (anthropic model settings changed)

* fix: make kevlar have default reasoning on

* fix: bump anthropic sdk version

* fix: patch name

* pin newer version anthropic

---------

Co-authored-by: Ari Webb <ari@letta.com>

fern/openapi.json

@@ -19789,6 +19789,19 @@
             ],
             "title": "Verbosity",
             "description": "Soft control for how verbose model output should be, used for GPT-5 models."
+          },
+          "effort": {
+            "anyOf": [
+              {
+                "type": "string",
+                "enum": ["low", "medium", "high"]
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Effort",
+            "description": "Effort level for Opus 4.5 model (controls token conservation). Not setting this gives similar performance to 'high'."
           }
         },
         "type": "object",
@@ -28905,6 +28918,19 @@
             "description": "Configurable thinking budget for extended thinking. Used for enable_reasoner and also for Google Vertex models like Gemini 2.5 Flash. Minimum value is 1024 when used with enable_reasoner.",
             "default": 0
           },
+          "effort": {
+            "anyOf": [
+              {
+                "type": "string",
+                "enum": ["low", "medium", "high"]
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Effort",
+            "description": "The effort level for Anthropic Opus 4.5 model (controls token spending). Not setting this gives similar performance to 'high'."
+          },
           "frequency_penalty": {
             "anyOf": [
               {
@@ -31069,6 +31095,19 @@
             "default": 0,
             "deprecated": true
           },
+          "effort": {
+            "anyOf": [
+              {
+                "type": "string",
+                "enum": ["low", "medium", "high"]
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Effort",
+            "description": "The effort level for Anthropic Opus 4.5 model (controls token spending). Not setting this gives similar performance to 'high'."
+          },
           "frequency_penalty": {
             "anyOf": [
               {

letta/llm_api/anthropic_client.py

@@ -72,6 +72,14 @@ class AnthropicClient(LLMClientBase):
         except Exception:
             pass
 
+        # Opus 4.5 effort parameter - to extend to other models, modify the model check
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.effort is not None:
+            betas.append("effort-2025-11-24")
+
+        # Context management for Opus 4.5 to preserve thinking blocks (improves cache hits)
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.enable_reasoner:
+            betas.append("context-management-2025-06-27")
+
         if betas:
             response = client.beta.messages.create(**request_data, betas=betas)
         else:
@@ -98,6 +106,14 @@ class AnthropicClient(LLMClientBase):
         except Exception:
             pass
 
+        # Opus 4.5 effort parameter - to extend to other models, modify the model check
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.effort is not None:
+            betas.append("effort-2025-11-24")
+
+        # Context management for Opus 4.5 to preserve thinking blocks (improves cache hits)
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.enable_reasoner:
+            betas.append("context-management-2025-06-27")
+
         if betas:
             response = await client.beta.messages.create(**request_data, betas=betas)
         else:
@@ -131,6 +147,14 @@ class AnthropicClient(LLMClientBase):
         except Exception:
             pass
 
+        # Opus 4.5 effort parameter - to extend to other models, modify the model check
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.effort is not None:
+            betas.append("effort-2025-11-24")
+
+        # Context management for Opus 4.5 to preserve thinking blocks (improves cache hits)
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.enable_reasoner:
+            betas.append("context-management-2025-06-27")
+
         return await client.beta.messages.create(**request_data, betas=betas)
 
     @trace_method
@@ -271,6 +295,23 @@ class AnthropicClient(LLMClientBase):
             # Silently disable prefix_fill for now
             prefix_fill = False
 
+        # Effort configuration for Opus 4.5 (controls token spending)
+        # To extend to other models, modify the model check
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.effort is not None:
+            data["output_config"] = {"effort": llm_config.effort}
+
+        # Context management for Opus 4.5 to preserve thinking blocks and improve cache hits
+        # See: https://docs.anthropic.com/en/docs/build-with-claude/context-editing
+        if llm_config.model.startswith("claude-opus-4-5") and llm_config.enable_reasoner:
+            data["context_management"] = {
+                "edits": [
+                    {
+                        "type": "clear_thinking_20251015",
+                        "keep": "all",  # Preserve all thinking blocks for maximum cache performance
+                    }
+                ]
+            }
+
         # Tools
         # For an overview on tool choice:
         # https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview
@@ -541,6 +582,17 @@ class AnthropicClient(LLMClientBase):
             except Exception:
                 pass
 
+            # Opus 4.5 beta flags for effort and context management
+            # Note: effort beta is added if model is kevlar (actual effort value is in count_params)
+            # Context management beta is added for consistency with main requests
+            if model and model.startswith("claude-opus-4-5"):
+                # Add effort beta if output_config is present in count_params
+                if "output_config" in count_params:
+                    betas.append("effort-2025-11-24")
+                # Add context management beta if thinking is enabled
+                if thinking_enabled:
+                    betas.append("context-management-2025-06-27")
+
             if betas:
                 result = await client.beta.messages.count_tokens(**count_params, betas=betas)
             else:
@@ -559,6 +611,8 @@ class AnthropicClient(LLMClientBase):
             or llm_config.model.startswith("claude-sonnet-4")
             or llm_config.model.startswith("claude-opus-4")
             or llm_config.model.startswith("claude-haiku-4-5")
+            # Opus 4.5 support - to extend effort parameter to other models, modify this check
+            or llm_config.model.startswith("claude-opus-4-5")
         )
 
     @trace_method

letta/schemas/llm_config.py

@@ -77,6 +77,10 @@ class LLMConfig(BaseModel):
         0,
         description="Configurable thinking budget for extended thinking. Used for enable_reasoner and also for Google Vertex models like Gemini 2.5 Flash. Minimum value is 1024 when used with enable_reasoner.",
     )
+    effort: Optional[Literal["low", "medium", "high"]] = Field(
+        None,
+        description="The effort level for Anthropic Opus 4.5 model (controls token spending). Not setting this gives similar performance to 'high'.",
+    )
     frequency_penalty: Optional[float] = Field(
         None,  # Can also deafult to 0.0?
         description="Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim. From OpenAI: Number between -2.0 and 2.0.",
@@ -172,6 +176,7 @@ class LLMConfig(BaseModel):
             or model.startswith("claude-sonnet-4")
             or model.startswith("claude-opus-4")
             or model.startswith("claude-haiku-4-5")
+            or model.startswith("claude-opus-4-5")
         ):
             values["put_inner_thoughts_in_kwargs"] = False
 
@@ -372,6 +377,7 @@ class LLMConfig(BaseModel):
             or config.model.startswith("claude-sonnet-4")
             or config.model.startswith("claude-3-7-sonnet")
             or config.model.startswith("claude-haiku-4-5")
+            or config.model.startswith("claude-opus-4-5")
         )
 
     @classmethod

letta/schemas/model.py

@@ -268,6 +268,12 @@ class AnthropicModelSettings(ModelSettings):
         description="Soft control for how verbose model output should be, used for GPT-5 models.",
     )
 
+    # Opus 4.5 effort parameter
+    effort: Optional[Literal["low", "medium", "high"]] = Field(
+        None,
+        description="Effort level for Opus 4.5 model (controls token conservation). Not setting this gives similar performance to 'high'.",
+    )
+
     # TODO: implement support for these
     # top_k: Optional[int] = Field(None, description="The number of top tokens to return.")
     # top_p: Optional[float] = Field(None, description="The top-p value to use when generating text.")
@@ -280,6 +286,7 @@ class AnthropicModelSettings(ModelSettings):
             "thinking_budget_tokens": self.thinking.budget_tokens,
             "verbosity": self.verbosity,
             "parallel_tool_calls": self.parallel_tool_calls,
+            "effort": self.effort,
         }
 
 

letta/schemas/providers/anthropic.py

@@ -93,6 +93,11 @@ MODEL_LIST = [
         "name": "claude-3-5-haiku-latest",
         "context_window": 200000,
     },
+    ## Opus 4.5
+    {
+        "name": "claude-opus-4-5-20251101",
+        "context_window": 200000,
+    },
 ]
 
 

pyproject.toml

@@ -42,7 +42,7 @@ dependencies = [
     "grpcio-tools>=1.68.1",
     "llama-index>=0.12.2",
     "llama-index-embeddings-openai>=0.3.1",
-    "anthropic>=0.49.0",
+    "anthropic>=0.75.0",
     "letta-client>=0.1.319",
     "openai>=1.99.9",
     "opentelemetry-api==1.30.0",

uv.lock

@@ -209,20 +209,21 @@ wheels = [
 
 [[package]]
 name = "anthropic"
-version = "0.64.0"
+version = "0.75.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "anyio" },
     { name = "distro" },
+    { name = "docstring-parser" },
     { name = "httpx" },
     { name = "jiter" },
     { name = "pydantic" },
     { name = "sniffio" },
     { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/d8/4f/f2b880cba1a76f3acc7d5eb2ae217632eac1b8cef5ed3027493545c59eba/anthropic-0.64.0.tar.gz", hash = "sha256:3d496c91a63dff64f451b3e8e4b238a9640bf87b0c11d0b74ddc372ba5a3fe58", size = 427893, upload-time = "2025-08-13T17:09:49.915Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/04/1f/08e95f4b7e2d35205ae5dcbb4ae97e7d477fc521c275c02609e2931ece2d/anthropic-0.75.0.tar.gz", hash = "sha256:e8607422f4ab616db2ea5baacc215dd5f028da99ce2f022e33c7c535b29f3dfb", size = 439565, upload-time = "2025-11-24T20:41:45.28Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/a9/b2/2d268bcd5d6441df9dc0ebebc67107657edb8b0150d3fda1a5b81d1bec45/anthropic-0.64.0-py3-none-any.whl", hash = "sha256:6f5f7d913a6a95eb7f8e1bda4e75f76670e8acd8d4cd965e02e2a256b0429dd1", size = 297244, upload-time = "2025-08-13T17:09:47.908Z" },
+    { url = "https://files.pythonhosted.org/packages/60/1c/1cd02b7ae64302a6e06724bf80a96401d5313708651d277b1458504a1730/anthropic-0.75.0-py3-none-any.whl", hash = "sha256:ea8317271b6c15d80225a9f3c670152746e88805a7a61e14d4a374577164965b", size = 388164, upload-time = "2025-11-24T20:41:43.587Z" },
 ]
 
 [[package]]
@@ -2486,7 +2487,7 @@ requires-dist = [
     { name = "aiosqlite", marker = "extra == 'desktop'", specifier = ">=0.21.0" },
     { name = "aiosqlite", marker = "extra == 'sqlite'", specifier = ">=0.21.0" },
     { name = "alembic", specifier = ">=1.13.3" },
-    { name = "anthropic", specifier = ">=0.49.0" },
+    { name = "anthropic", specifier = ">=0.75.0" },
     { name = "apscheduler", specifier = ">=3.11.0" },
     { name = "async-lru", marker = "extra == 'desktop'", specifier = ">=2.0.5" },
     { name = "asyncpg", marker = "extra == 'postgres'", specifier = ">=0.30.0" },