diff --git a/CHANGELOG.md b/CHANGELOG.md
index f49cd574..cf877128 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -86,6 +86,11 @@ APIs change between minors without a deprecation period.
 - A picture in a prompt was counted as its base64 by the context check, so
   one 1280x720 PNG was "138 thousand tokens" and a vision model refused it;
   a picture or a sound now counts as about a thousand.
+- An on-demand model of a higher class was evicted, idle, for a lower-class
+  one two minutes into a lull; it keeps the card through its idle timeout,
+  and a model waiting on it is told so.
+- A model removed from the catalog stayed wanted, so added again it loaded
+  unasked.
 
 
 ## v0.14.3 — 2026-10-10
diff --git a/internal/control/reconciler/evict.go b/internal/control/reconciler/evict.go
index 212607ab..0bbbb15d 100644
--- a/internal/control/reconciler/evict.go
+++ b/internal/control/reconciler/evict.go
@@ -90,6 +90,11 @@ func (r *Reconciler) evictable(wantedBy string) []evictionCandidate {
 	cat := r.cat.Load()
 	now := r.now()
 
+	wantRank := -1
+	if w, ok := cat.Resolve(wantedBy); ok {
+		wantRank = catalog.ClassRank(w.EffectiveClass())
+	}
+
 	r.mu.Lock()
 	defer r.mu.Unlock()
 
@@ -102,6 +107,12 @@ func (r *Reconciler) evictable(wantedBy string) []evictionCandidate {
 		if !ok || !other.Scaling.OnDemand() {
 			continue
 		}
+		// Not for a model of a lower class: the one put first keeps the
+		// card through its idle timeout, rather than losing it two minutes
+		// into a lull to whatever was asked for next.
+		if catalog.ClassRank(other.EffectiveClass()) > wantRank {
+			continue
+		}
 		// Unknown is not idle. A replica the router has said nothing about —
 		// after a restart, that is all of them — must not be evicted on the
 		// strength of a fact nobody has reported.
diff --git a/internal/control/reconciler/health.go b/internal/control/reconciler/health.go
index 38ea71d3..dbdb5dc6 100644
--- a/internal/control/reconciler/health.go
+++ b/internal/control/reconciler/health.go
@@ -201,13 +201,17 @@ func (r *Reconciler) WaitFor(model string) (heldBy string, eta time.Duration, ok
 		}
 		h := holder{v: victim{ID: rep.ID, Model: rep.Model}}
 		if !rep.LastUsedAt.IsZero() {
-			evict := rep.LastUsedAt.Add(r.cfg.EvictionGrace)
-			if !rep.ReadyAt.IsZero() && rep.ReadyAt.Add(r.cfg.MinLifetime).After(evict) {
-				evict = rep.ReadyAt.Add(r.cfg.MinLifetime)
-			}
-			h.release = evict
-			if idle := rep.LastUsedAt.Add(r.idleTimeout(other)); idle.Before(evict) {
-				h.release = idle
+			h.release = rep.LastUsedAt.Add(r.idleTimeout(other))
+			// Evicted sooner, when it is no higher a class than the model
+			// waiting; one put first is waited for to its idle timeout.
+			if catalog.ClassRank(other.EffectiveClass()) <= catalog.ClassRank(m.EffectiveClass()) {
+				evict := rep.LastUsedAt.Add(r.cfg.EvictionGrace)
+				if !rep.ReadyAt.IsZero() && rep.ReadyAt.Add(r.cfg.MinLifetime).After(evict) {
+					evict = rep.ReadyAt.Add(r.cfg.MinLifetime)
+				}
+				if evict.Before(h.release) {
+					h.release = evict
+				}
 			}
 		}
 		holders = append(holders, h)
diff --git a/internal/control/reconciler/reconciler_test.go b/internal/control/reconciler/reconciler_test.go
index 0467828d..56ef40f2 100644
--- a/internal/control/reconciler/reconciler_test.go
+++ b/internal/control/reconciler/reconciler_test.go
@@ -1337,6 +1337,48 @@ func TestAModelRemovedWhileRunningIsStopped(t *testing.T) {
 	}
 }
 
+func TestAModelPutFirstIsNotEvictedForALowerClassOne(t *testing.T) {
+	// The VLM is put first: a picture or 3D model asked for in a lull must
+	// wait for its idle timeout, not take the card two minutes in.
+	h := newHarness(t, testModel(1), gpuNode("1", 4))
+	cat := twoOnDemandModels(t)
+	ms := cat.Models()
+	for i := range ms {
+		if ms[i].Name == "llama" {
+			ms[i].Class = catalog.ClassStandard
+		} else {
+			ms[i].Class = catalog.ClassOpportunistic
+		}
+	}
+	built, err := catalog.New(ms)
+	if err != nil {
+		t.Fatal(err)
+	}
+	h.r.cat.Store(built)
+	h.r.Want("llama")
+	h.r.Reconcile(t.Context())
+	h.markReady()
+	reps := h.r.Replicas()
+	h.r.RecordUsage(map[string]time.Time{reps[0].ID: h.clock.now()})
+	h.advance(DefaultConfig().MinLifetime + DefaultConfig().EvictionGrace + time.Minute)
+
+	if got := h.r.evictable("second"); len(got) != 0 {
+		t.Errorf("a standard model was offered to an opportunistic one: %+v", got)
+	}
+	h.r.Want("second")
+	h.r.Reconcile(t.Context())
+	// It waits for the idle timeout (an hour here), and is told so.
+	_, eta, ok := h.r.WaitFor("second")
+	want := time.Hour - (DefaultConfig().MinLifetime + DefaultConfig().EvictionGrace + time.Minute)
+	if !ok || eta != want {
+		t.Errorf("WaitFor: %s %v, want %s", eta, ok, want)
+	}
+	// The other way, it is offered.
+	if got := h.r.evictable("llama"); len(got) != 0 {
+		t.Errorf("nothing of second runs, yet %+v was offered", got)
+	}
+}
+
 func TestAnOnDemandModelRemovedAndAddedAgainIsNotLoadedUnasked(t *testing.T) {
 	// The VLM, removed and added again to give it a licence, came back
 	// loading on its own: still on the wanted list from before its removal.