Skip to content

Automate browser based workflows with AI

Pin GPT-6 Luna to high reasoning effort, matching Task V3's GPT-5.6 Luna (SKY-16907) (#8669)

Shuchang Zhengb325eb10b04f
diff --git a/skyvern/config.py b/skyvern/config.py
index 087bbe2d0..e2856be21 100644
--- a/skyvern/config.py
+++ b/skyvern/config.py
@@ -517,6 +517,8 @@ class Settings(BaseSettings):
# OPENAI
OPENAI_API_KEY: str | None = None
GPT5_REASONING_EFFORT: str | None = "medium"
+ # Pinned on every GPT-6 Luna key; high matches the gpt-5.6-luna key Task V3 serves (OPENAI_GPT5_6_LUNA_HIGH).
+ GPT6_LUNA_REASONING_EFFORT: str = "high"
OPENAI_CUA_MODEL: str = "computer-use-preview"
# xAI
XAI_API_KEY: str | None = None
diff --git a/skyvern/forge/sdk/api/llm/config_registry.py b/skyvern/forge/sdk/api/llm/config_registry.py
index fc1763c8f..855d36baa 100644
--- a/skyvern/forge/sdk/api/llm/config_registry.py
+++ b/skyvern/forge/sdk/api/llm/config_registry.py
@@ -14,6 +14,11 @@ from skyvern.schemas.llm import LiteLLMParams, LLMConfig, LLMRouterConfig
LOG = structlog.get_logger()
FLEX_EXECUTION_TIMEOUT_SECONDS = 180.0
+GPT6_REASONING_EFFORT: dict[str, str] = {
+ "astra": "xhigh",
+ "sol": "xhigh",
+ "luna": settings.GPT6_LUNA_REASONING_EFFORT,
+}
XAI_GROK_4_5_MODEL = "xai/grok-4.5"
XAI_GROK_4_5_CONTEXT_WINDOW = 500_000
# xAI publishes no output cap for grok-4.5; match the bound used by the other large reasoning
@@ -374,7 +379,7 @@ if settings.ENABLE_OPENAI:
reasoning_effort=settings.GPT5_REASONING_EFFORT,
),
)
- for variant in ("astra", "sol", "luna"):
+ for variant, gpt6_effort in GPT6_REASONING_EFFORT.items():
LLMConfigRegistry.register_config(
f"OPENAI_GPT6_{variant.upper()}",
LLMConfig(
@@ -384,7 +389,7 @@ if settings.ENABLE_OPENAI:
add_assistant_prefix=False,
max_completion_tokens=128000,
temperature=None,
- reasoning_effort="xhigh",
+ reasoning_effort=gpt6_effort,
pin_reasoning_effort=True,
),
)
@@ -1132,7 +1137,7 @@ if settings.ENABLE_AZURE_GPT5_4:
),
)
-for variant in ("astra", "sol", "luna"):
+for variant, gpt6_effort in GPT6_REASONING_EFFORT.items():
prefix = f"AZURE_GPT6_{variant.upper()}"
if getattr(settings, f"ENABLE_{prefix}"):
LLMConfigRegistry.register_config(
@@ -1150,7 +1155,7 @@ for variant in ("astra", "sol", "luna"):
add_assistant_prefix=False,
max_completion_tokens=128000,
temperature=None,
- reasoning_effort="xhigh",
+ reasoning_effort=gpt6_effort,
pin_reasoning_effort=True,
),
)