diff --git a/plugins/model-providers/upstage/__init__.py b/plugins/model-providers/upstage/__init__.py
index d36c0734a3a9..b050051badef 100644
--- a/plugins/model-providers/upstage/__init__.py
+++ b/plugins/model-providers/upstage/__init__.py
@@ -76,18 +76,20 @@ class UpstageProfile(ProviderProfile):
             return {}, top_level
 
         # Map Hermes' effort vocabulary onto Solar's accepted set. xhigh/max/
-        # ultra collapse to high (Solar's strongest). minimal → off (omit). An
-        # enabled request with no recognised effort uses the default effort.
+        # ultra collapse to high (Solar's strongest). minimal → off (omit).
+        # Unknown-but-enabled efforts (future vocabulary additions above
+        # "high", per the max/ultra precedent in #62650) also collapse to
+        # high rather than silently downgrading to the medium default.
         effort = (reasoning_config.get("effort") or "").strip().lower()
+        if not effort:
+            top_level["reasoning_effort"] = _DEFAULT_REASONING_EFFORT
+            return {}, top_level
         mapped = {
             "minimal": None,
             "low": "low",
             "medium": "medium",
             "high": "high",
-            "xhigh": "high",
-            "max": "high",
-            "ultra": "high",
-        }.get(effort, _DEFAULT_REASONING_EFFORT)
+        }.get(effort, "high")
 
         if mapped:
             top_level["reasoning_effort"] = mapped
diff --git a/tests/hermes_cli/test_upstage_provider.py b/tests/hermes_cli/test_upstage_provider.py
index 4e2829f61419..6f7da10761f7 100644
--- a/tests/hermes_cli/test_upstage_provider.py
+++ b/tests/hermes_cli/test_upstage_provider.py
@@ -12,7 +12,6 @@ from __future__ import annotations
 import sys
 import types
 
-import pytest
 
 if "dotenv" not in sys.modules:
     fake_dotenv = types.ModuleType("dotenv")
diff --git a/tests/plugins/model_providers/test_upstage_profile.py b/tests/plugins/model_providers/test_upstage_profile.py
index 257900aed198..ce227b675d8f 100644
--- a/tests/plugins/model_providers/test_upstage_profile.py
+++ b/tests/plugins/model_providers/test_upstage_profile.py
@@ -49,9 +49,12 @@ class TestUpstageProfile:
         # Only the agentic, tool-calling Solar Pro models belong in the offline
         # catalog — Mini is capable but not agentic, so it's never promoted as a
         # default. Live /v1/models still surfaces everything when a key is set.
-        assert upstage_profile.fallback_models == (
-            "solar-pro3",
-        )
+        # Behavior contract (not a frozen list): non-empty, no denied families.
+        assert upstage_profile.fallback_models
+        for denied in ("solar-mini", "syn-pro"):
+            assert not any(
+                denied in m for m in upstage_profile.fallback_models
+            ), f"non-agentic family {denied!r} must not be a fallback default"
 
     def test_default_model_is_solar_pro3(self, upstage_profile):
         # Entry [0] is the setup default (get_default_model_for_provider).
@@ -85,6 +88,16 @@ class TestUpstageReasoning:
         )
         assert top_level == {"reasoning_effort": "high"}
 
+    def test_unknown_future_effort_collapses_to_high(self, upstage_profile):
+        # Guard against the #62650 recurrence: a future effort level Hermes
+        # adds above "high" must collapse to Solar's strongest, not silently
+        # downgrade to the medium default.
+        _, top_level = upstage_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "hyperthink"},
+            model="solar-pro3",
+        )
+        assert top_level == {"reasoning_effort": "high"}
+
     def test_pro_enabled_without_effort_defaults_on(self, upstage_profile):
         _, top_level = upstage_profile.build_api_kwargs_extras(
             reasoning_config={"enabled": True}, model="solar-pro3"