Fimeg/letta-server
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

feat: add Anthropic Opus 4.6 model support (#9123)

Browse Source
This commit is contained in:
Kevin Lin
2026-02-05 10:23:59 -08:00
committed by Caren Thomas
parent ff69c6a32e
commit 34159ffa21
4 changed files with 84 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fern/openapi.json
View File

@@ -36592,14 +36592,14 @@
"anyOf": [
{
"type": "string",
"enum": ["low", "medium", "high"]
"enum": ["low", "medium", "high", "max"]
},
{
"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'."
"description": "The effort level for Anthropic models that support it (Opus 4.5, Opus 4.6). Controls token spending and thinking behavior. Not setting this gives similar performance to 'high'."
},
"frequency_penalty": {
"anyOf": [
@@ -39118,14 +39118,14 @@
"anyOf": [
{
"type": "string",
"enum": ["low", "medium", "high"]
"enum": ["low", "medium", "high", "max"]
},
{
"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'."
"description": "The effort level for Anthropic models that support it (Opus 4.5, Opus 4.6). Controls token spending and thinking behavior. Not setting this gives similar performance to 'high'."
},
"frequency_penalty": {
"anyOf": [

64
letta/llm_api/anthropic_client.py
View File

@@ -62,9 +62,15 @@ class AnthropicClient(LLMClientBase):
def request(self, request_data: dict, llm_config: LLMConfig) -> dict:
client = self._get_anthropic_client(llm_config, async_client=False)
betas: list[str] = []
# Interleaved thinking for reasoner (sync path parity)
# Opus 4.6 Auto Thinking
if llm_config.enable_reasoner:
if llm_config.model.startswith("claude-opus-4-6"):
betas.append("adaptive-thinking-2026-01-28")
# Interleaved thinking for other reasoners (sync path parity)
else:
betas.append("interleaved-thinking-2025-05-14")
# 1M context beta for Sonnet 4/4.5 when enabled
try:
from letta.settings import model_settings
@@ -76,9 +82,14 @@ 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:
# Effort parameter for Opus 4.5 and Opus 4.6 - to extend to other models, modify the model check
if (
llm_config.model.startswith("claude-opus-4-5") or llm_config.model.startswith("claude-opus-4-6")
) and llm_config.effort is not None:
betas.append("effort-2025-11-24")
# Max effort beta for Opus 4.6
if llm_config.model.startswith("claude-opus-4-6") and llm_config.effort == "max":
betas.append("max-effort-2026-01-24")
# 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:
@@ -98,8 +109,13 @@ class AnthropicClient(LLMClientBase):
async def request_async(self, request_data: dict, llm_config: LLMConfig) -> dict:
client = await self._get_anthropic_client_async(llm_config, async_client=True)
betas: list[str] = []
# interleaved thinking for reasoner
# Opus 4.6 Auto Thinking
if llm_config.enable_reasoner:
if llm_config.model.startswith("claude-opus-4-6"):
betas.append("adaptive-thinking-2026-01-28")
# Interleaved thinking for other reasoners (sync path parity)
else:
betas.append("interleaved-thinking-2025-05-14")
# 1M context beta for Sonnet 4/4.5 when enabled
@@ -113,9 +129,14 @@ 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:
# Effort parameter for Opus 4.5 and Opus 4.6 - to extend to other models, modify the model check
if (
llm_config.model.startswith("claude-opus-4-5") or llm_config.model.startswith("claude-opus-4-6")
) and llm_config.effort is not None:
betas.append("effort-2025-11-24")
# Max effort beta for Opus 4.6
if llm_config.model.startswith("claude-opus-4-6") and llm_config.effort == "max":
betas.append("max-effort-2026-01-24")
# 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:
@@ -262,9 +283,12 @@ class AnthropicClient(LLMClientBase):
# See: https://docs.anthropic.com/en/docs/build-with-claude/tool-use/fine-grained-streaming
betas = ["fine-grained-tool-streaming-2025-05-14"]
# If extended thinking, turn on interleaved header
# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#interleaved-thinking
# Opus 4.6 Auto Thinking
if llm_config.enable_reasoner:
if llm_config.model.startswith("claude-opus-4-6"):
betas.append("adaptive-thinking-2026-01-28")
# Interleaved thinking for other reasoners (sync path parity)
else:
betas.append("interleaved-thinking-2025-05-14")
# 1M context beta for Sonnet 4/4.5 when enabled
@@ -278,9 +302,14 @@ 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:
# Effort parameter for Opus 4.5 and Opus 4.6 - to extend to other models, modify the model check
if (
llm_config.model.startswith("claude-opus-4-5") or llm_config.model.startswith("claude-opus-4-6")
) and llm_config.effort is not None:
betas.append("effort-2025-11-24")
# Max effort beta for Opus 4.6
if llm_config.model.startswith("claude-opus-4-6") and llm_config.effort == "max":
betas.append("max-effort-2026-01-24")
# 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:
@@ -462,6 +491,13 @@ class AnthropicClient(LLMClientBase):
# Extended Thinking
if self.is_reasoning_model(llm_config) and llm_config.enable_reasoner:
# Opus 4.6 uses Auto Thinking (no budget tokens)
if llm_config.model.startswith("claude-opus-4-6"):
data["thinking"] = {
"type": "adaptive",
}
else:
# Traditional extended thinking with budget tokens
thinking_budget = max(llm_config.max_reasoning_tokens, 1024)
if thinking_budget != llm_config.max_reasoning_tokens:
logger.warning(
@@ -477,9 +513,11 @@ class AnthropicClient(LLMClientBase):
# Silently disable prefix_fill for now
prefix_fill = False
# Effort configuration for Opus 4.5 (controls token spending)
# Effort configuration for Opus 4.5 and Opus 4.6 (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:
if (
llm_config.model.startswith("claude-opus-4-5") or llm_config.model.startswith("claude-opus-4-6")
) 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
@@ -851,6 +889,8 @@ class AnthropicClient(LLMClientBase):
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")
# Opus 4.6 support - uses Auto Thinking
or llm_config.model.startswith("claude-opus-4-6")
)
@trace_method

14
letta/schemas/llm_config.py
View File

@@ -82,9 +82,9 @@ 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(
effort: Optional[Literal["low", "medium", "high", "max"]] = Field(
None,
description="The effort level for Anthropic Opus 4.5 model (controls token spending). Not setting this gives similar performance to 'high'.",
description="The effort level for Anthropic models that support it (Opus 4.5, Opus 4.6). Controls token spending and thinking behavior. Not setting this gives similar performance to 'high'.",
)
frequency_penalty: Optional[float] = Field(
None, # Can also deafult to 0.0?
@@ -190,6 +190,7 @@ class LLMConfig(BaseModel):
or model.startswith("claude-opus-4")
or model.startswith("claude-haiku-4-5")
or model.startswith("claude-opus-4-5")
or model.startswith("claude-opus-4-6")
):
values["put_inner_thoughts_in_kwargs"] = False
@@ -441,6 +442,7 @@ class LLMConfig(BaseModel):
or config.model.startswith("claude-3-7-sonnet")
or config.model.startswith("claude-haiku-4-5")
or config.model.startswith("claude-opus-4-5")
or config.model.startswith("claude-opus-4-6")
)
@classmethod
@@ -543,8 +545,8 @@ class LLMConfig(BaseModel):
config.put_inner_thoughts_in_kwargs = False
if config.enable_reasoner and config.max_reasoning_tokens == 0:
config.max_reasoning_tokens = 1024
# Set default effort level for Claude Opus 4.5
if config.model.startswith("claude-opus-4-5") and config.effort is None:
# Set default effort level for Claude Opus 4.5 and Opus 4.6
if (config.model.startswith("claude-opus-4-5") or config.model.startswith("claude-opus-4-6")) and config.effort is None:
config.effort = "medium"
return config
@@ -612,8 +614,8 @@ class LLMConfig(BaseModel):
config.put_inner_thoughts_in_kwargs = False
if config.max_reasoning_tokens == 0:
config.max_reasoning_tokens = 1024
# Set default effort level for Claude Opus 4.5
if config.model.startswith("claude-opus-4-5") and config.effort is None:
# Set default effort level for Claude Opus 4.5 and Opus 4.6
if (config.model.startswith("claude-opus-4-5") or config.model.startswith("claude-opus-4-6")) and config.effort is None:
config.effort = "medium"
elif cls.is_google_vertex_reasoning_model(config) or cls.is_google_ai_reasoning_model(config):
# Handle as non-reasoner until we support summary

9
letta/schemas/providers/anthropic.py
View File

@@ -108,6 +108,11 @@ MODEL_LIST = [
"name": "claude-opus-4-5-20251101",
"context_window": 200000,
},
## Opus 4.6
{
"name": "claude-opus-4-6",
"context_window": 200000,
},
]
@@ -134,7 +139,9 @@ class AnthropicProvider(Provider):
def get_default_max_output_tokens(self, model_name: str) -> int:
"""Get the default max output tokens for Anthropic models."""
if "opus" in model_name:
if "claude-opus-4-6" in model_name:
return 21000 # Opus 4.6 supports up to 128k with streaming, use 21k as default
elif "opus" in model_name:
return 16384
elif "sonnet" in model_name:
return 16384