diff --git a/.changeset/fortunate-friendly-avocet.md b/.changeset/fortunate-friendly-avocet.md
new file mode 100644
index 00000000..5f1ee443
--- /dev/null
+++ b/.changeset/fortunate-friendly-avocet.md
@@ -0,0 +1,5 @@
+---
+"stagehand": patch
+---
+
+Fix ability to pass raw JSON to Extract schema
diff --git a/stagehand/handlers/extract_handler.py b/stagehand/handlers/extract_handler.py
index 8af621c9..5fa24212 100644
--- a/stagehand/handlers/extract_handler.py
+++ b/stagehand/handlers/extract_handler.py
@@ -151,9 +151,7 @@ async def extract(
 
         processed_data_payload = raw_data_dict  # Default to the raw dictionary
 
-        if schema and isinstance(
-            raw_data_dict, dict
-        ):  # schema is the Pydantic model type
+        if schema and isinstance(schema, type) and issubclass(schema, BaseModel):
             # Try direct validation first
             try:
                 validated_model_instance = schema.model_validate(raw_data_dict)
diff --git a/stagehand/llm/inference.py b/stagehand/llm/inference.py
index b438883b..5c9d122c 100644
--- a/stagehand/llm/inference.py
+++ b/stagehand/llm/inference.py
@@ -169,11 +169,19 @@ async def extract(
     start_time = time.time()
 
     # Determine if we need to use schema-based response format
-    # TODO: if schema is json, return json
     response_format = {"type": "json_object"}
     if schema:
-        # If schema is a Pydantic model, use it directly
-        response_format = schema
+        if isinstance(schema, dict):
+            response_format = {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": "extraction_schema",
+                    "strict": False,
+                    "schema": schema,
+                },
+            }
+        else:
+            response_format = schema
 
     # Call the LLM with appropriate parameters
     try: