diff --git a/src/tabpfn_client/api_models.py b/src/tabpfn_client/api_models.py
index cf1a5a8..941e0d8 100644
--- a/src/tabpfn_client/api_models.py
+++ b/src/tabpfn_client/api_models.py
@@ -108,7 +108,7 @@ class FitRequest(BaseModel):
     # Budget for the fit (seconds). Only consulted when `thinking_effort` is set.
     thinking_timeout_s: Optional[float] = None
     # Optimization metric for the fit. Only consulted when `thinking_effort` is set.
-    thinking_metric: Optional[str] = None
+    thinking_effort_metric: Optional[str] = None
 
 
 class FitResponse(BaseModel):
diff --git a/src/tabpfn_client/client.py b/src/tabpfn_client/client.py
index 8ab3a4c..170ed4c 100644
--- a/src/tabpfn_client/client.py
+++ b/src/tabpfn_client/client.py
@@ -407,6 +407,8 @@ def fit(
         # The client-side `thinking_*` knobs forward 1:1 to the server's
         # top-level FitRequest fields. When the user enabled thinking via
         # `thinking_mode=True` without picking a level, default to "medium".
+        # The user-facing kwarg is `thinking_metric`; on the wire it is sent
+        # as `thinking_effort_metric` (matching the server's FitRequest schema).
         if thinking_enabled and tabpfn_config:
             thinking_effort = tabpfn_config.get("thinking_effort") or "medium"
             thinking_timeout_s = tabpfn_config.get("thinking_timeout_s")
@@ -444,7 +446,7 @@ def fit(
                 tabpfn_config=server_tabpfn_config,
                 thinking_effort=thinking_effort,
                 thinking_timeout_s=thinking_timeout_s,
-                thinking_metric=thinking_metric,
+                thinking_effort_metric=thinking_metric,
             ),
             timeout=client_options.timeout,
             headers=client_options.headers,