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

feat: add Anthropic Sonnet 4.6 (#9408)

Browse Source
This commit is contained in:
Devansh Jain
2026-02-17 10:13:05 -08:00
committed by Caren Thomas
parent 25d54dd896
commit 39ddda81cc
4 changed files with 56 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
letta/llm_api/anthropic_client.py
View File

@@ -65,9 +65,9 @@ class AnthropicClient(LLMClientBase):
client = self._get_anthropic_client(llm_config, async_client=False)
betas: list[str] = []
# Opus 4.6 Auto Thinking
# Opus 4.6 / Sonnet 4.6 Auto Thinking
if llm_config.enable_reasoner:
if llm_config.model.startswith("claude-opus-4-6"):
if llm_config.model.startswith("claude-opus-4-6") or llm_config.model.startswith("claude-sonnet-4-6"):
betas.append("adaptive-thinking-2026-01-28")
# Interleaved thinking for other reasoners (sync path parity)
else:
@@ -86,13 +86,17 @@ class AnthropicClient(LLMClientBase):
except Exception:
pass
# Effort parameter for Opus 4.5 and Opus 4.6 - to extend to other models, modify the model check
# Effort parameter for Opus 4.5, Opus 4.6, and Sonnet 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")
llm_config.model.startswith("claude-opus-4-5")
or llm_config.model.startswith("claude-opus-4-6")
or llm_config.model.startswith("claude-sonnet-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":
# Max effort beta for Opus 4.6 / Sonnet 4.6
if (
llm_config.model.startswith("claude-opus-4-6") or llm_config.model.startswith("claude-sonnet-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)
@@ -134,9 +138,9 @@ class AnthropicClient(LLMClientBase):
client = await self._get_anthropic_client_async(llm_config, async_client=True)
betas: list[str] = []
# Opus 4.6 Auto Thinking
# Opus 4.6 / Sonnet 4.6 Auto Thinking
if llm_config.enable_reasoner:
if llm_config.model.startswith("claude-opus-4-6"):
if llm_config.model.startswith("claude-opus-4-6") or llm_config.model.startswith("claude-sonnet-4-6"):
betas.append("adaptive-thinking-2026-01-28")
# Interleaved thinking for other reasoners (sync path parity)
else:
@@ -155,13 +159,17 @@ class AnthropicClient(LLMClientBase):
except Exception:
pass
# Effort parameter for Opus 4.5 and Opus 4.6 - to extend to other models, modify the model check
# Effort parameter for Opus 4.5, Opus 4.6, and Sonnet 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")
llm_config.model.startswith("claude-opus-4-5")
or llm_config.model.startswith("claude-opus-4-6")
or llm_config.model.startswith("claude-sonnet-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":
# Max effort beta for Opus 4.6 / Sonnet 4.6
if (
llm_config.model.startswith("claude-opus-4-6") or llm_config.model.startswith("claude-sonnet-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)
@@ -311,9 +319,9 @@ 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"]
# Opus 4.6 Auto Thinking
# Opus 4.6 / Sonnet 4.6 Auto Thinking
if llm_config.enable_reasoner:
if llm_config.model.startswith("claude-opus-4-6"):
if llm_config.model.startswith("claude-opus-4-6") or llm_config.model.startswith("claude-sonnet-4-6"):
betas.append("adaptive-thinking-2026-01-28")
# Interleaved thinking for other reasoners (sync path parity)
else:
@@ -332,13 +340,17 @@ class AnthropicClient(LLMClientBase):
except Exception:
pass
# Effort parameter for Opus 4.5 and Opus 4.6 - to extend to other models, modify the model check
# Effort parameter for Opus 4.5, Opus 4.6, and Sonnet 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")
llm_config.model.startswith("claude-opus-4-5")
or llm_config.model.startswith("claude-opus-4-6")
or llm_config.model.startswith("claude-sonnet-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":
# Max effort beta for Opus 4.6 / Sonnet 4.6
if (
llm_config.model.startswith("claude-opus-4-6") or llm_config.model.startswith("claude-sonnet-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)
@@ -528,8 +540,8 @@ class AnthropicClient(LLMClientBase):
)
if should_enable_thinking:
# Opus 4.6 uses Auto Thinking (no budget tokens)
if llm_config.model.startswith("claude-opus-4-6"):
# Opus 4.6 / Sonnet 4.6 uses Auto Thinking (no budget tokens)
if llm_config.model.startswith("claude-opus-4-6") or llm_config.model.startswith("claude-sonnet-4-6"):
data["thinking"] = {
"type": "adaptive",
}
@@ -550,10 +562,12 @@ class AnthropicClient(LLMClientBase):
# Silently disable prefix_fill for now
prefix_fill = False
# Effort configuration for Opus 4.5 and Opus 4.6 (controls token spending)
# Effort configuration for Opus 4.5, Opus 4.6, and Sonnet 4.6 (controls token spending)
# 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")
llm_config.model.startswith("claude-opus-4-5")
or llm_config.model.startswith("claude-opus-4-6")
or llm_config.model.startswith("claude-sonnet-4-6")
) and llm_config.effort is not None:
data["output_config"] = {"effort": llm_config.effort}
@@ -935,6 +949,8 @@ class AnthropicClient(LLMClientBase):
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")
# Sonnet 4.6 support - same API as Opus 4.6
or llm_config.model.startswith("claude-sonnet-4-6")
)
@trace_method

12
letta/schemas/llm_config.py
View File

@@ -562,7 +562,11 @@ class LLMConfig(BaseModel):
if config.enable_reasoner and config.max_reasoning_tokens == 0:
config.max_reasoning_tokens = 1024
# 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:
if (
config.model.startswith("claude-opus-4-5")
or config.model.startswith("claude-opus-4-6")
or config.model.startswith("claude-sonnet-4-6")
) and config.effort is None:
config.effort = "medium"
return config
@@ -631,7 +635,11 @@ class LLMConfig(BaseModel):
if config.max_reasoning_tokens == 0:
config.max_reasoning_tokens = 1024
# 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:
if (
config.model.startswith("claude-opus-4-5")
or config.model.startswith("claude-opus-4-6")
or config.model.startswith("claude-sonnet-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

@@ -113,6 +113,11 @@ MODEL_LIST = [
"name": "claude-opus-4-6",
"context_window": 200000,
},
## Sonnet 4.6
{
"name": "claude-sonnet-4-6",
"context_window": 200000,
},
]
@@ -139,8 +144,8 @@ class AnthropicProvider(Provider):
def get_default_max_output_tokens(self, model_name: str) -> int:
"""Get the default max output tokens for Anthropic models."""
if "claude-opus-4-6" in model_name:
return 21000 # Opus 4.6 supports up to 128k with streaming, use 21k as default
if "claude-opus-4-6" in model_name or "claude-sonnet-4-6" in model_name:
return 21000 # Opus 4.6 / Sonnet 4.6 supports up to 128k with streaming, use 21k as default
elif "opus" in model_name:
return 16384
elif "sonnet" in model_name:

2
letta/settings.py
View File

@@ -170,7 +170,7 @@ class ModelSettings(BaseSettings):
anthropic_sonnet_1m: bool = Field(
default=False,
description=(
"Enable 1M-token context window for Claude Sonnet 4/4.5. When true, adds the"
"Enable 1M-token context window for Claude Sonnet 4/4.5/4.6. When true, adds the"
" 'context-1m-2025-08-07' beta to Anthropic requests and sets model context_window"
" to 1,000,000 instead of 200,000. Note: This feature is in beta and not available"
" to all orgs; once GA, this flag can be removed and behavior can default to on."