diff --git a/internal/proxy/service.go b/internal/proxy/service.go
index 623699481..f14875b8e 100644
--- a/internal/proxy/service.go
+++ b/internal/proxy/service.go
@@ -2283,12 +2283,13 @@ func (s *Service) ProxyMessages(ctx context.Context, body []byte, w http.Respons
 	// Semantic-cache eligibility: configured, non-streaming, decision has
 	// metadata, externalID present, not eval traffic. Skip when a compaction
 	// handover rewrote env (embedding predates the rewrite) or when subsidy
-	// factors are non-empty (the cache key doesn't capture quota-headroom-
-	// dependent model choice; subsidyFactors returns nil when the feature is off).
-	// Subscription-only turns are excluded (like the OpenAI path): the mode is an
-	// unfoldable routing signal absent from the cache key, so a stored body would
-	// bypass the exhausted-sub 402 guard and the depleted-credits warning below.
-	cacheEligible := s.semanticCache != nil && !env.Stream() && decision.Metadata != nil && externalID != "" && !bypassEval && !compactionHandoverRan && !billing.SubscriptionOnlyFromContext(ctx) && len(s.subsidyFactors(ctx, r.Header)) == 0
+	// factors / PreferredModels are non-empty (the cache key doesn't capture
+	// those score-perturbing inputs' effect on model choice; subsidyFactors
+	// returns nil when the feature is off). Subscription-only turns are
+	// excluded (like the OpenAI path): the mode is an unfoldable routing signal
+	// absent from the cache key, so a stored body would bypass the exhausted-sub
+	// 402 guard and the depleted-credits warning below.
+	cacheEligible := s.semanticCache != nil && !env.Stream() && decision.Metadata != nil && externalID != "" && !bypassEval && !compactionHandoverRan && !billing.SubscriptionOnlyFromContext(ctx) && len(s.subsidyFactors(ctx, r.Header)) == 0 && len(s.preferredModelsForRequest(ctx)) == 0
 	if cacheEligible {
 		if resp, hit := s.semanticCache.Lookup(externalID, cache.FormatAnthropic, decision.Metadata.Embedding, decision.Metadata.ClusterIDs, decision.Metadata.ClusterRouterVersion, decision.Metadata.EffectiveKnobsHash); hit {
 			s.writeCachedResponse(w, resp, decision)
@@ -4213,9 +4214,10 @@ func (s *Service) ProxyOpenAIChatCompletion(ctx context.Context, body []byte, w
 	pinAgeSec := routeRes.PinAgeSec
 	s.logPlannerOutcome(ctx, routeRes)
 
-	// See the ProxyMessages cache-eligibility note: subsidized requests bypass the
-	// semantic cache (the key doesn't capture headroom-dependent model choice).
-	cacheEligible := s.semanticCache != nil && !env.Stream() && decision.Metadata != nil && externalID != "" && !bypassEval && !responsesPassthrough && !billing.SubscriptionOnlyFromContext(ctx) && len(s.subsidyFactors(ctx, r.Header)) == 0
+	// See the ProxyMessages cache-eligibility note: subsidized and
+	// PreferredModels requests bypass the semantic cache (the key doesn't
+	// capture those score-perturbing inputs' effect on model choice).
+	cacheEligible := s.semanticCache != nil && !env.Stream() && decision.Metadata != nil && externalID != "" && !bypassEval && !responsesPassthrough && !billing.SubscriptionOnlyFromContext(ctx) && len(s.subsidyFactors(ctx, r.Header)) == 0 && len(s.preferredModelsForRequest(ctx)) == 0
 	if cacheEligible {
 		if resp, hit := s.semanticCache.Lookup(externalID, cache.FormatOpenAI, decision.Metadata.Embedding, decision.Metadata.ClusterIDs, decision.Metadata.ClusterRouterVersion, decision.Metadata.EffectiveKnobsHash); hit {
 			s.writeCachedResponse(w, resp, decision)
diff --git a/internal/proxy/service_cache_test.go b/internal/proxy/service_cache_test.go
index f57fe0520..64dc6f5b1 100644
--- a/internal/proxy/service_cache_test.go
+++ b/internal/proxy/service_cache_test.go
@@ -237,3 +237,40 @@ func TestService_Cache_DisabledByNilCache(t *testing.T) {
 	assert.Len(t, provider.proxyBodies, 2, "nil cache must be a transparent passthrough")
 	assert.Empty(t, rec2.Header().Get(proxy.HeaderRouterCache))
 }
+
+// TestService_Cache_PreferredModelsBypasses guards #789: PreferredModels is a
+// per-request score-perturbing input (blendScoresV2 priorityBonus) that can
+// flip the scorer's winner without changing EffectiveKnobsHash. Without a
+// cacheEligible gate, a no-preference store is replayed to a preference-
+// bearing request (wrong model's body, live x-router-model). Same class as
+// the subsidyFactors gate from #495.
+func TestService_Cache_PreferredModelsBypasses(t *testing.T) {
+	emb := embeddingFixture(8)
+	provider := &fakeProvider{
+		proxyResponse: func(w http.ResponseWriter) {
+			w.Header().Set("Content-Type", "application/json")
+			_, _ = w.Write([]byte(`{"id":"from-no-pref","content":"opus-shaped"}`))
+		},
+	}
+	// Fake router returns a fixed decision+embedding so both turns land in the
+	// same cache bucket — isolating the eligibility gate from scorer variance.
+	fr := &fakeRouter{decision: decisionWithEmbedding(emb, []int{0, 1, 2, 3})}
+	c := cache.New(cache.DefaultConfig())
+	svc := proxy.NewService(fr, map[string]providers.Client{providers.ProviderAnthropic: provider}, nil, false, c, nil, false, providers.ProviderAnthropic, "claude-haiku-4-5", nil)
+
+	body := anthropicBody("close-call prompt", false)
+
+	ctxNoPref := proxyContextWithExternalID(t, "tenant-pref")
+	rec1 := httptest.NewRecorder()
+	require.NoError(t, svc.ProxyMessages(ctxNoPref, body, rec1, httptest.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(""))))
+	require.Len(t, provider.proxyBodies, 1, "no-preference turn must miss and hit the provider")
+
+	ctxPref := context.WithValue(ctxNoPref, proxy.InstallationPreferredModelsContextKey{}, []string{"claude-haiku-4-5"})
+	rec2 := httptest.NewRecorder()
+	require.NoError(t, svc.ProxyMessages(ctxPref, body, rec2, httptest.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(""))))
+
+	assert.Len(t, provider.proxyBodies, 2,
+		"PreferredModels turn must bypass the semantic cache and hit the provider (no cross-preference hit)")
+	assert.Empty(t, rec2.Header().Get(proxy.HeaderRouterCache),
+		"PreferredModels turn must not report x-router-cache: hit")
+}