diff --git a/internal/affinity/key.go b/internal/affinity/key.go
index 0fc0e40d..aebdc0a3 100644
--- a/internal/affinity/key.go
+++ b/internal/affinity/key.go
@@ -7,7 +7,7 @@ import (
 	"gpt-load/internal/protocol"
 )
 
-const keyDomain = "gpt-load/affinity/prompt-prefix-hmac/v1"
+const keyDomain = "gpt-load/affinity/prompt-prefix-hmac/v2"
 
 // Key is an opaque, tenant-scoped affinity cache key.
 type Key string
@@ -26,7 +26,6 @@ func DeriveKey(
 	hasher Hasher,
 	accessKeyID uint,
 	clientProtocol protocol.Protocol,
-	externalModel string,
 	prefix []byte,
 ) Key {
 	if hasher == nil || accessKeyID == 0 || !clientProtocol.Valid() || len(prefix) == 0 {
@@ -38,7 +37,6 @@ func DeriveKey(
 	binary.BigEndian.PutUint64(encodedID[:], uint64(accessKeyID))
 	material.Write(encodedID[:])
 	writeKeyField(&material, []byte(clientProtocol))
-	writeKeyField(&material, []byte(externalModel))
 	writeKeyField(&material, prefix)
 	return Key(hasher.Hash(material.String()))
 }
diff --git a/internal/affinity/key_test.go b/internal/affinity/key_test.go
index 4a30222a..88406a8a 100644
--- a/internal/affinity/key_test.go
+++ b/internal/affinity/key_test.go
@@ -18,7 +18,7 @@ func (testHasher) Hash(value string) string {
 func TestDeriveKeyScopesStablePrefix(t *testing.T) {
 	t.Parallel()
 
-	base := DeriveKey(testHasher{}, 7, protocol.OpenAICompletions, "gpt-4o", []byte("stable-prefix"))
+	base := DeriveKey(testHasher{}, 7, protocol.OpenAICompletions, []byte("stable-prefix"))
 	if !base.Valid() {
 		t.Fatal("DeriveKey() returned an empty key")
 	}
@@ -26,23 +26,21 @@ func TestDeriveKeyScopesStablePrefix(t *testing.T) {
 		name      string
 		accessKey uint
 		protocol  protocol.Protocol
-		model     string
 		prefix    string
 	}{
-		{name: "access key", accessKey: 8, protocol: protocol.OpenAICompletions, model: "gpt-4o", prefix: "stable-prefix"},
-		{name: "protocol", accessKey: 7, protocol: protocol.Anthropic, model: "gpt-4o", prefix: "stable-prefix"},
-		{name: "model", accessKey: 7, protocol: protocol.OpenAICompletions, model: "gpt-4o-mini", prefix: "stable-prefix"},
-		{name: "prefix", accessKey: 7, protocol: protocol.OpenAICompletions, model: "gpt-4o", prefix: "other-prefix"},
+		{name: "access key", accessKey: 8, protocol: protocol.OpenAICompletions, prefix: "stable-prefix"},
+		{name: "protocol", accessKey: 7, protocol: protocol.Anthropic, prefix: "stable-prefix"},
+		{name: "prefix", accessKey: 7, protocol: protocol.OpenAICompletions, prefix: "other-prefix"},
 	}
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
-			got := DeriveKey(testHasher{}, test.accessKey, test.protocol, test.model, []byte(test.prefix))
+			got := DeriveKey(testHasher{}, test.accessKey, test.protocol, []byte(test.prefix))
 			if !got.Valid() || got == base {
 				t.Fatalf("DeriveKey() = %q, want non-empty key different from base", got)
 			}
 		})
 	}
-	if duplicate := DeriveKey(testHasher{}, 7, protocol.OpenAICompletions, "gpt-4o", []byte("stable-prefix")); duplicate != base {
+	if duplicate := DeriveKey(testHasher{}, 7, protocol.OpenAICompletions, []byte("stable-prefix")); duplicate != base {
 		t.Fatalf("duplicate DeriveKey() = %q, want %q", duplicate, base)
 	}
 }
@@ -64,7 +62,7 @@ func TestDeriveKeyRejectsIncompleteScope(t *testing.T) {
 	}
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
-			if got := DeriveKey(test.hasher, test.accessKey, test.protocol, "model", test.prefix); got.Valid() {
+			if got := DeriveKey(test.hasher, test.accessKey, test.protocol, test.prefix); got.Valid() {
 				t.Fatalf("DeriveKey() = %q, want empty key", got)
 			}
 		})
@@ -74,8 +72,8 @@ func TestDeriveKeyRejectsIncompleteScope(t *testing.T) {
 func TestDeriveKeyUsesUnambiguousFieldBoundaries(t *testing.T) {
 	t.Parallel()
 
-	left := DeriveKey(testHasher{}, 1, protocol.OpenAICompletions, "ab", []byte("c"))
-	right := DeriveKey(testHasher{}, 1, protocol.OpenAICompletions, "a", []byte("bc"))
+	left := DeriveKey(testHasher{}, 1, protocol.OpenAICompletions, []byte("ab"))
+	right := DeriveKey(testHasher{}, 1, protocol.OpenAICompletions, []byte("a"))
 	if !left.Valid() || !right.Valid() || left == right {
 		t.Fatalf("keys = %q / %q, want distinct non-empty values", left, right)
 	}
diff --git a/internal/gateway/affinity.go b/internal/gateway/affinity.go
index 3bcb8678..af9ae74c 100644
--- a/internal/gateway/affinity.go
+++ b/internal/gateway/affinity.go
@@ -17,7 +17,6 @@ func (handler *Handler) resolveRequestAffinity(
 	snapshot *state.ConfigSnapshot,
 	accessKeyID uint,
 	clientProtocol protocol.Protocol,
-	externalModel string,
 	prefix []byte,
 	allowedCredentialRefs map[uint]state.CredentialRef,
 ) requestAffinity {
@@ -33,7 +32,6 @@ func (handler *Handler) resolveRequestAffinity(
 		handler.encryption,
 		accessKeyID,
 		clientProtocol,
-		externalModel,
 		prefix,
 	)
 	if !key.Valid() {
diff --git a/internal/gateway/affinity_test.go b/internal/gateway/affinity_test.go
index 52ae24fa..62a72bd3 100644
--- a/internal/gateway/affinity_test.go
+++ b/internal/gateway/affinity_test.go
@@ -12,6 +12,7 @@ import (
 	"github.com/gin-gonic/gin"
 
 	"gpt-load/internal/affinity"
+	"gpt-load/internal/execution"
 	"gpt-load/internal/platform/config"
 	"gpt-load/internal/protocol"
 	"gpt-load/internal/state"
@@ -56,6 +57,41 @@ func TestHandlerLearnsAndReusesAutomaticSoftAffinity(t *testing.T) {
 	assertAffinityHits(t, sink.snapshot(), []bool{false, true})
 }
 
+func TestHandlerReusesSoftAffinityAcrossModelsWithSameCandidateRange(t *testing.T) {
+	forwarder := &scriptedForwarder{results: successfulAffinityResults(2)}
+	handler, manager, _ := newHandlerForTest(t, forwarder, "sk-one", "sk-two")
+	addAffinityModelRoute(t, manager.Current(), "gpt-4o-mini", 1)
+	sink := &recordingRequestLogSink{}
+	handler.requestLogSink = sink
+	useAffinityRandomValues(handler, 0, affinitySecondCredentialRand)
+	engine := newAffinityTestEngine(t, handler)
+
+	serveAffinityModelRequest(t, engine, "gpt-4o")
+	serveAffinityModelRequest(t, engine, "gpt-4o-mini")
+
+	assertAffinityAttemptKeys(t, forwarder.inputs, []string{"sk-one", "sk-one"})
+	assertAffinityHits(t, sink.snapshot(), []bool{false, true})
+}
+
+func TestHandlerRelearnsSoftAffinityWhenModelCandidateRangeChanges(t *testing.T) {
+	forwarder := &scriptedForwarder{results: successfulAffinityResults(4)}
+	handler, manager, registry := newHandlerForTest(t, forwarder, "sk-one", "sk-two")
+	moveSecondAffinityCredentialToGroup(t, manager.Current(), registry)
+	sink := &recordingRequestLogSink{}
+	handler.requestLogSink = sink
+	useAffinityRandomValues(handler, 0, 0, 0, 0)
+	engine := newAffinityTestEngine(t, handler)
+
+	serveAffinityModelRequest(t, engine, "gpt-4o")
+	serveAffinityModelRequest(t, engine, "gpt-4o-mini")
+	serveAffinityModelRequest(t, engine, "gpt-4o-mini")
+	serveAffinityModelRequest(t, engine, "gpt-4o")
+
+	assertAffinityAttemptKeys(t, forwarder.inputs, []string{"sk-one", "sk-two", "sk-two", "sk-one"})
+	assertAffinityHits(t, sink.snapshot(), []bool{false, false, true, false})
+	assertAffinityUpstreamModels(t, forwarder.inputs, []string{"gpt-4o", "gpt-4o-mini", "gpt-4o-mini", "gpt-4o"})
+}
+
 func TestHandlerDoesNotLearnAffinityForNonParticipatingGroup(t *testing.T) {
 	forwarder := &scriptedForwarder{results: successfulAffinityResults(2)}
 	handler, manager, _ := newHandlerForTest(t, forwarder, "sk-one", "sk-two")
@@ -227,7 +263,6 @@ func TestHandlerIgnoresAffinityAfterCredentialIdentityChanges(t *testing.T) {
 		snapshot,
 		1,
 		protocol.OpenAICompletions,
-		"gpt-4o",
 		prefix,
 		map[uint]state.CredentialRef{1: oldRef},
 	)
@@ -246,7 +281,6 @@ func TestHandlerIgnoresAffinityAfterCredentialIdentityChanges(t *testing.T) {
 		snapshot,
 		1,
 		protocol.OpenAICompletions,
-		"gpt-4o",
 		prefix,
 		map[uint]state.CredentialRef{1: oldRef},
 	)
@@ -259,7 +293,6 @@ func TestHandlerIgnoresAffinityAfterCredentialIdentityChanges(t *testing.T) {
 		snapshot,
 		1,
 		protocol.OpenAICompletions,
-		"gpt-4o",
 		prefix,
 		map[uint]state.CredentialRef{1: changedRef},
 	)
@@ -297,6 +330,75 @@ func useAffinityRandomValues(handler *Handler, values ...int64) {
 	}
 }
 
+func addAffinityModelRoute(
+	t *testing.T,
+	snapshot *state.ConfigSnapshot,
+	externalModel string,
+	groupID uint,
+) {
+	t.Helper()
+	byOperation := snapshot.ExecutionCandidates[protocol.OpenAICompletions]
+	if byOperation == nil {
+		t.Fatal("OpenAI Completions execution candidates are missing")
+	}
+	byModel := byOperation[execution.OperationChatCompletion]
+	base := byModel["gpt-4o"]
+	if len(base) != 1 {
+		t.Fatalf("gpt-4o routes = %d, want 1", len(base))
+	}
+	route := base[0]
+	route.GroupID = groupID
+	route.UpstreamModelID = externalModel
+	byModel[externalModel] = []state.RouteTarget{route}
+}
+
+func moveSecondAffinityCredentialToGroup(
+	t *testing.T,
+	snapshot *state.ConfigSnapshot,
+	registry *state.CredentialRegistry,
+) {
+	t.Helper()
+	group := snapshot.Groups[1]
+	group.ID = 2
+	group.Name = "openai-two"
+	group.Models = []state.ModelConfig{{ID: "gpt-4o-mini"}}
+	snapshot.Groups[2] = group
+	catalog := snapshot.GroupCatalog[1]
+	catalog.ID = 2
+	catalog.Name = group.Name
+	snapshot.GroupCatalog[2] = catalog
+	addAffinityModelRoute(t, snapshot, "gpt-4o-mini", 2)
+
+	refs := registry.CaptureActiveCredentialRefs([]uint{1})
+	if len(refs) != 2 {
+		t.Fatalf("captured credential refs = %d, want 2", len(refs))
+	}
+	entries := make([]state.CredentialEntry, 0, len(refs))
+	for _, ref := range refs {
+		groupID := uint(1)
+		if ref.ID == 2 {
+			groupID = 2
+		}
+		entries = append(entries, state.CredentialEntry{
+			ID: ref.ID, GroupID: groupID, Version: ref.Version,
+			IdentityGeneration: ref.IdentityGeneration, Fingerprint: ref.Fingerprint,
+			Status: state.CredentialStatusActive, EncryptedValue: ref.EncryptedValue,
+		})
+	}
+	if err := registry.ReplaceCredentials(entries); err != nil {
+		t.Fatalf("ReplaceCredentials() error = %v", err)
+	}
+}
+
+func serveAffinityModelRequest(t *testing.T, engine http.Handler, model string) {
+	t.Helper()
+	serveAffinityRequest(
+		t,
+		engine,
+		`{"model":"`+model+`","messages":[{"role":"user","content":"stable conversation"}]}`,
+	)
+}
+
 func newAffinityTestEngine(t *testing.T, handler *Handler) http.Handler {
 	t.Helper()
 	engine := gin.New()
@@ -340,3 +442,14 @@ func assertAffinityHits(t *testing.T, events []telemetry.RequestEvent, want []bo
 		t.Fatalf("affinity hits = %#v, want %#v", got, want)
 	}
 }
+
+func assertAffinityUpstreamModels(t *testing.T, inputs []ForwardInput, want []string) {
+	t.Helper()
+	got := make([]string, 0, len(inputs))
+	for _, input := range inputs {
+		got = append(got, input.UpstreamModelID)
+	}
+	if !reflect.DeepEqual(got, want) {
+		t.Fatalf("upstream models = %#v, want %#v", got, want)
+	}
+}
diff --git a/internal/gateway/handler.go b/internal/gateway/handler.go
index a408a1f8..a87a4452 100644
--- a/internal/gateway/handler.go
+++ b/internal/gateway/handler.go
@@ -444,7 +444,6 @@ func (handler *Handler) Handle(ginContext *gin.Context) {
 		snapshot,
 		accessKey.ID,
 		selectedRoute.Protocol,
-		model,
 		metadata.AffinityPrefix,
 		allowedCredentialRefs,
 	)