fix(proxy)!: enforce fallback budget by default

A budget bypass that ships off by default stays open for every deployment
that does not know to look for the flag, so `enforce_fallback_budget` now
defaults to true and `general_settings.enforce_fallback_budget: false` is
the opt-out for anyone who wants the old unguarded behaviour back.

BREAKING CHANGE: a paid fallback target is now refused for callers who are
over their key or user `max_budget`. Deployments relying on fallbacks to
keep serving over-budget callers must set enforce_fallback_budget: false.

litellm/proxy/auth/fallback_budget.py

@@ -8,8 +8,8 @@ actually bills. So a free model with a paid fallback spends without a gate.
 
 This predicate is injected into the router to re-check budget for each fallback target before it is
 attempted, mirroring `fallback_model_access.py`. It deliberately leaves the primary attempt alone:
-a zero-cost model is never blocked by budget, and only the paid fallback is refused. Opt-in via
-`general_settings.enforce_fallback_budget: true`.
+a zero-cost model is never blocked by budget, and only the paid fallback is refused. On by default;
+set `general_settings.enforce_fallback_budget: false` to restore the unguarded behaviour.
 
 Scope: the key's and the user's `max_budget`. Not covered yet, and each needs a read-only evaluation
 path before it can be: team, team-member, end-user, org, global and per-model budgets, whose
@@ -50,7 +50,7 @@ class _RequestMetadata(BaseModel):
 
 
 class _FallbackBudgetSettings(BaseModel):
-    enforce_fallback_budget: bool = False
+    enforce_fallback_budget: bool = True
 
 
 def _token_in_metadata(metadata: object) -> UserAPIKeyAuth | None:

tests/test_litellm/proxy/auth/test_fallback_budget.py

@@ -5,6 +5,7 @@ from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.auth.fallback_budget import (
     RouterFallbackBudgetCheck,
     is_token_within_budget_for_model,
+    router_fallback_budget_check,
 )
 
 FREE_MODEL = {
@@ -182,3 +183,20 @@ async def test_router_without_a_budget_check_attempts_every_fallback():
     over = {"metadata": {"user_api_key_auth": _token(user_spend=1900.0, user_max_budget=50.0)}}
 
     assert await _is_fallback_target_within_budget(router, "paid-model", "free-model", over) is True
+
+
+@pytest.mark.asyncio
+async def test_enforcement_is_on_by_default_and_opt_out_restores_the_leak(monkeypatch):
+    """
+    Leaving the paid fallback unguarded is the budget bypass this module exists to close, so an
+    unconfigured proxy has to enforce. `enforce_fallback_budget: false` is the deliberate opt-out.
+    """
+    from litellm.proxy import proxy_server
+
+    over = {"metadata": {"user_api_key_auth": _token(user_spend=1900.0, user_max_budget=50.0)}}
+
+    monkeypatch.setattr(proxy_server, "general_settings", {}, raising=False)
+    assert await router_fallback_budget_check(model="paid-model", request_kwargs=over, llm_router=_router()) is False
+
+    monkeypatch.setattr(proxy_server, "general_settings", {"enforce_fallback_budget": False}, raising=False)
+    assert await router_fallback_budget_check(model="paid-model", request_kwargs=over, llm_router=_router()) is True

tests/test_litellm/proxy/test_proxy_server.py

@@ -13416,14 +13416,15 @@ async def test_load_config_router_budget_checks_fallback_targets_against_the_cal
         }
     }
 
-    # off by default: the paid fallback is still attempted for an over-budget caller
+    # on by default: an over-budget caller is refused the paid fallback with no config at all
     monkeypatch.setattr(proxy_server, "general_settings", {}, raising=False)
-    assert await router.fallback_budget_check(model="m", request_kwargs=over_budget, llm_router=router) is True
-
-    monkeypatch.setattr(proxy_server, "general_settings", {"enforce_fallback_budget": True}, raising=False)
     assert await router.fallback_budget_check(model="m", request_kwargs=over_budget, llm_router=router) is False
     assert await router.fallback_budget_check(model="m", request_kwargs=under_budget, llm_router=router) is True
 
+    # explicit opt-out restores the unguarded behaviour
+    monkeypatch.setattr(proxy_server, "general_settings", {"enforce_fallback_budget": False}, raising=False)
+    assert await router.fallback_budget_check(model="m", request_kwargs=over_budget, llm_router=router) is True
+
 
 @pytest.mark.asyncio
 async def test_load_config_user_api_key_cache_max_size_keeps_more_than_200_entries(tmp_path, monkeypatch):