diff --git a/py/autoevals/ragas.py b/py/autoevals/ragas.py
index 4726159..04430f5 100644
--- a/py/autoevals/ragas.py
+++ b/py/autoevals/ragas.py
@@ -115,7 +115,6 @@ def context_relevancy_scorer(output, expected, input, metadata):
 from .llm import OpenAILLMScorer
 from .oai import (
     Client,
-    _default_model_var,
     arun_cached_request,
     get_default_embedding_model,
     get_default_model,
@@ -130,28 +129,16 @@ def check_required(name, **kwargs):
             raise ValueError(f"{name} requires {key} value")
 
 
-# Deprecated: Use init(default_model="...") to configure the default model instead.
-# This was previously "gpt-5-nano" but now defaults to the configured model.
-DEFAULT_RAGAS_MODEL = "gpt-5-nano"
-
-
 def _get_model(model: str | None) -> str:
     """Get the model to use, respecting init(default_model=...) configuration.
 
-    Falls back to DEFAULT_RAGAS_MODEL if no model is specified and no custom
-    default has been configured.
+    Falls back to the library-wide default when no model is specified and no
+    custom default has been configured.
     """
     if model is not None:
         return model
 
-    # Check if user configured a custom default via init(default_model=...)
-    # If they did (even if it's "gpt-5-mini"), respect it for consistency
-    configured_default = _default_model_var.get(None)
-    if configured_default is not None:
-        return configured_default
-
-    # Fall back to RAGAS-specific default when user hasn't configured anything
-    return DEFAULT_RAGAS_MODEL
+    return get_default_model()
 
 
 ENTITY_PROMPT = """Given a text, extract unique entities without repetition. Ensure you consider different forms or mentions of the same entity as a single entity.
diff --git a/py/autoevals/test_ragas_default_model.py b/py/autoevals/test_ragas_default_model.py
new file mode 100644
index 0000000..c292c1a
--- /dev/null
+++ b/py/autoevals/test_ragas_default_model.py
@@ -0,0 +1,29 @@
+"""The RAGAS scorers must not fall back to their own hardcoded model.
+
+`_get_model` is what every RAGAS scorer calls to resolve its judge model.
+Before this change it returned `DEFAULT_RAGAS_MODEL` ("gpt-5-nano") when the
+caller passed no model and had not called `init(default_model=...)`, so a
+RAGAS scorer silently used a different judge than every other scorer in the
+library -- and than the TypeScript implementation, which resolves through
+`getDefaultModel()`. The comment above the constant already said the
+hardcoded fallback was deprecated.
+"""
+
+from autoevals import init
+from autoevals.oai import get_default_model
+from autoevals.ragas import _get_model
+
+
+def test_ragas_default_matches_library_default():
+    init()
+    assert _get_model(None) == get_default_model()
+
+
+def test_ragas_respects_explicit_model():
+    init()
+    assert _get_model("gpt-4o") == "gpt-4o"
+
+
+def test_ragas_respects_configured_default():
+    init(default_model="gpt-4-turbo")
+    assert _get_model(None) == "gpt-4-turbo"