diff --git a/backend/internal/service/openai_compat_prompt_cache_key.go b/backend/internal/service/openai_compat_prompt_cache_key.go
new file mode 100644
index 00000000..88e16a4d
--- /dev/null
+++ b/backend/internal/service/openai_compat_prompt_cache_key.go
@@ -0,0 +1,81 @@
+package service
+
+import (
+	"encoding/json"
+	"strings"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/apicompat"
+)
+
+const compatPromptCacheKeyPrefix = "compat_cc_"
+
+func shouldAutoInjectPromptCacheKeyForCompat(model string) bool {
+	switch normalizeCodexModel(strings.TrimSpace(model)) {
+	case "gpt-5.4", "gpt-5.3-codex":
+		return true
+	default:
+		return false
+	}
+}
+
+func deriveCompatPromptCacheKey(req *apicompat.ChatCompletionsRequest, mappedModel string) string {
+	if req == nil {
+		return ""
+	}
+
+	normalizedModel := normalizeCodexModel(strings.TrimSpace(mappedModel))
+	if normalizedModel == "" {
+		normalizedModel = normalizeCodexModel(strings.TrimSpace(req.Model))
+	}
+	if normalizedModel == "" {
+		normalizedModel = strings.TrimSpace(req.Model)
+	}
+
+	seedParts := []string{"model=" + normalizedModel}
+	if req.ReasoningEffort != "" {
+		seedParts = append(seedParts, "reasoning_effort="+strings.TrimSpace(req.ReasoningEffort))
+	}
+	if len(req.ToolChoice) > 0 {
+		seedParts = append(seedParts, "tool_choice="+normalizeCompatSeedJSON(req.ToolChoice))
+	}
+	if len(req.Tools) > 0 {
+		if raw, err := json.Marshal(req.Tools); err == nil {
+			seedParts = append(seedParts, "tools="+normalizeCompatSeedJSON(raw))
+		}
+	}
+	if len(req.Functions) > 0 {
+		if raw, err := json.Marshal(req.Functions); err == nil {
+			seedParts = append(seedParts, "functions="+normalizeCompatSeedJSON(raw))
+		}
+	}
+
+	firstUserCaptured := false
+	for _, msg := range req.Messages {
+		switch strings.TrimSpace(msg.Role) {
+		case "system":
+			seedParts = append(seedParts, "system="+normalizeCompatSeedJSON(msg.Content))
+		case "user":
+			if !firstUserCaptured {
+				seedParts = append(seedParts, "first_user="+normalizeCompatSeedJSON(msg.Content))
+				firstUserCaptured = true
+			}
+		}
+	}
+
+	return compatPromptCacheKeyPrefix + hashSensitiveValueForLog(strings.Join(seedParts, "|"))
+}
+
+func normalizeCompatSeedJSON(v json.RawMessage) string {
+	if len(v) == 0 {
+		return ""
+	}
+	var tmp any
+	if err := json.Unmarshal(v, &tmp); err != nil {
+		return string(v)
+	}
+	out, err := json.Marshal(tmp)
+	if err != nil {
+		return string(v)
+	}
+	return string(out)
+}
diff --git a/backend/internal/service/openai_compat_prompt_cache_key_test.go b/backend/internal/service/openai_compat_prompt_cache_key_test.go
new file mode 100644
index 00000000..eb9148de
--- /dev/null
+++ b/backend/internal/service/openai_compat_prompt_cache_key_test.go
@@ -0,0 +1,64 @@
+package service
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/apicompat"
+	"github.com/stretchr/testify/require"
+)
+
+func mustRawJSON(t *testing.T, s string) json.RawMessage {
+	t.Helper()
+	return json.RawMessage(s)
+}
+
+func TestShouldAutoInjectPromptCacheKeyForCompat(t *testing.T) {
+	require.True(t, shouldAutoInjectPromptCacheKeyForCompat("gpt-5.4"))
+	require.True(t, shouldAutoInjectPromptCacheKeyForCompat("gpt-5.3"))
+	require.True(t, shouldAutoInjectPromptCacheKeyForCompat("gpt-5.3-codex"))
+	require.False(t, shouldAutoInjectPromptCacheKeyForCompat("gpt-4o"))
+}
+
+func TestDeriveCompatPromptCacheKey_StableAcrossLaterTurns(t *testing.T) {
+	base := &apicompat.ChatCompletionsRequest{
+		Model: "gpt-5.4",
+		Messages: []apicompat.ChatMessage{
+			{Role: "system", Content: mustRawJSON(t, `"You are helpful."`)},
+			{Role: "user", Content: mustRawJSON(t, `"Hello"`)},
+		},
+	}
+	extended := &apicompat.ChatCompletionsRequest{
+		Model: "gpt-5.4",
+		Messages: []apicompat.ChatMessage{
+			{Role: "system", Content: mustRawJSON(t, `"You are helpful."`)},
+			{Role: "user", Content: mustRawJSON(t, `"Hello"`)},
+			{Role: "assistant", Content: mustRawJSON(t, `"Hi there!"`)},
+			{Role: "user", Content: mustRawJSON(t, `"How are you?"`)},
+		},
+	}
+
+	k1 := deriveCompatPromptCacheKey(base, "gpt-5.4")
+	k2 := deriveCompatPromptCacheKey(extended, "gpt-5.4")
+	require.Equal(t, k1, k2, "cache key should be stable across later turns")
+	require.NotEmpty(t, k1)
+}
+
+func TestDeriveCompatPromptCacheKey_DiffersAcrossSessions(t *testing.T) {
+	req1 := &apicompat.ChatCompletionsRequest{
+		Model: "gpt-5.4",
+		Messages: []apicompat.ChatMessage{
+			{Role: "user", Content: mustRawJSON(t, `"Question A"`)},
+		},
+	}
+	req2 := &apicompat.ChatCompletionsRequest{
+		Model: "gpt-5.4",
+		Messages: []apicompat.ChatMessage{
+			{Role: "user", Content: mustRawJSON(t, `"Question B"`)},
+		},
+	}
+
+	k1 := deriveCompatPromptCacheKey(req1, "gpt-5.4")
+	k2 := deriveCompatPromptCacheKey(req2, "gpt-5.4")
+	require.NotEqual(t, k1, k2, "different first user messages should yield different keys")
+}
diff --git a/backend/internal/service/openai_gateway_chat_completions.go b/backend/internal/service/openai_gateway_chat_completions.go
index 7202f7cb..a442da33 100644
--- a/backend/internal/service/openai_gateway_chat_completions.go
+++ b/backend/internal/service/openai_gateway_chat_completions.go
@@ -43,23 +43,38 @@ func (s *OpenAIGatewayService) ForwardAsChatCompletions(
 	clientStream := chatReq.Stream
 	includeUsage := chatReq.StreamOptions != nil && chatReq.StreamOptions.IncludeUsage
 
-	// 2. Convert to Responses and forward
+	// 2. Resolve model mapping early so compat prompt_cache_key injection can
+	// derive a stable seed from the final upstream model family.
+	mappedModel := resolveOpenAIForwardModel(account, originalModel, defaultMappedModel)
+
+	promptCacheKey = strings.TrimSpace(promptCacheKey)
+	compatPromptCacheInjected := false
+	if promptCacheKey == "" && account.Type == AccountTypeOAuth && shouldAutoInjectPromptCacheKeyForCompat(mappedModel) {
+		promptCacheKey = deriveCompatPromptCacheKey(&chatReq, mappedModel)
+		compatPromptCacheInjected = promptCacheKey != ""
+	}
+
+	// 3. Convert to Responses and forward
 	// ChatCompletionsToResponses always sets Stream=true (upstream always streams).
 	responsesReq, err := apicompat.ChatCompletionsToResponses(&chatReq)
 	if err != nil {
 		return nil, fmt.Errorf("convert chat completions to responses: %w", err)
 	}
-
-	// 3. Model mapping
-	mappedModel := resolveOpenAIForwardModel(account, originalModel, defaultMappedModel)
 	responsesReq.Model = mappedModel
 
-	logger.L().Debug("openai chat_completions: model mapping applied",
+	logFields := []zap.Field{
 		zap.Int64("account_id", account.ID),
 		zap.String("original_model", originalModel),
 		zap.String("mapped_model", mappedModel),
 		zap.Bool("stream", clientStream),
-	)
+	}
+	if compatPromptCacheInjected {
+		logFields = append(logFields,
+			zap.Bool("compat_prompt_cache_key_injected", true),
+			zap.String("compat_prompt_cache_key_sha256", hashSensitiveValueForLog(promptCacheKey)),
+		)
+	}
+	logger.L().Debug("openai chat_completions: model mapping applied", logFields...)
 
 	// 4. Marshal Responses request body, then apply OAuth codex transform
 	responsesBody, err := json.Marshal(responsesReq)