diff --git a/backend/internal/handler/openai_images.go b/backend/internal/handler/openai_images.go
index d5d16d1e71..da41bf6aa3 100644
--- a/backend/internal/handler/openai_images.go
+++ b/backend/internal/handler/openai_images.go
@@ -144,7 +144,7 @@ func (h *OpenAIGatewayHandler) Images(c *gin.Context) {
 	}
 
 	sessionHash := h.gatewayService.GenerateExplicitSessionHash(c, body)
-	requestCtx := service.WithOpenAIImageGenerationIntent(c.Request.Context())
+	requestCtx := service.WithOpenAIImagesEndpoint(service.WithOpenAIImageGenerationIntent(c.Request.Context()))
 
 	maxAccountSwitches := h.maxAccountSwitches
 	switchCount := 0
diff --git a/backend/internal/pkg/ctxkey/ctxkey.go b/backend/internal/pkg/ctxkey/ctxkey.go
index 2ec3e0fdbe..c34ef589a0 100644
--- a/backend/internal/pkg/ctxkey/ctxkey.go
+++ b/backend/internal/pkg/ctxkey/ctxkey.go
@@ -50,6 +50,12 @@ const (
 	// OpenAIImageGenerationIntent 标识 OpenAI 请求会触发生图能力（用于图片能力维度限流）
 	OpenAIImageGenerationIntent Key = "ctx_openai_image_generation_intent"
 
+	// OpenAIImagesEndpoint 标识请求是从 /v1/images/* 入站的。
+	// 与 OpenAIImageGenerationIntent 的区别：后者只表示"这次请求会生图"，
+	// /v1/responses 带图片模型时也会置位；本 key 只在专用生图端点置位，
+	// 用于区分"用错端点"与"端点用对了但账号没能力"。
+	OpenAIImagesEndpoint Key = "ctx_openai_images_endpoint"
+
 	// Group 认证后的分组信息，由 API Key 认证中间件设置
 	Group Key = "ctx_group"
 
diff --git a/backend/internal/service/model_rate_limit.go b/backend/internal/service/model_rate_limit.go
index 195f5a1dfd..fe962b17c0 100644
--- a/backend/internal/service/model_rate_limit.go
+++ b/backend/internal/service/model_rate_limit.go
@@ -120,6 +120,23 @@ func OpenAIImageGenerationIntentFromContext(ctx context.Context) bool {
 	return ok && enabled
 }
 
+// WithOpenAIImagesEndpoint 标记请求从 /v1/images/* 专用生图端点入站。
+func WithOpenAIImagesEndpoint(ctx context.Context) context.Context {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, ctxkey.OpenAIImagesEndpoint, true)
+}
+
+// OpenAIImagesEndpointFromContext 报告请求是否来自 /v1/images/*。
+func OpenAIImagesEndpointFromContext(ctx context.Context) bool {
+	if ctx == nil {
+		return false
+	}
+	enabled, ok := ctx.Value(ctxkey.OpenAIImagesEndpoint).(bool)
+	return ok && enabled
+}
+
 func resolveFinalAntigravityModelKey(ctx context.Context, account *Account, requestedModel string) string {
 	modelKey := mapAntigravityModel(account, requestedModel)
 	if modelKey == "" {
diff --git a/backend/internal/service/ratelimit_service.go b/backend/internal/service/ratelimit_service.go
index ee75c28e48..335181dffa 100644
--- a/backend/internal/service/ratelimit_service.go
+++ b/backend/internal/service/ratelimit_service.go
@@ -2047,14 +2047,6 @@ func (s *RateLimitService) HandleUpstreamModelNotFound(ctx context.Context, acco
 	case isUpstreamModelNotFoundError(statusCode, responseBody):
 		cooldown, reason = upstreamModelNotFoundCooldown, upstreamModelNotFoundReason
 	case isOpenAIOAuthAccount(account) && isOpenAICodexPlanGatedModelError(statusCode, responseBody):
-		// Image models rejected by the Codex text endpoints are a deterministic
-		// endpoint mismatch, not an account capability gap: the same account still
-		// serves them over /v1/images/*. Cooling down the model here would take the
-		// whole pool offline for correct image requests, so fail the attempt over
-		// without recording a per-model cooldown.
-		if IsGPTImageGenerationModel(requestedModel) {
-			return true
-		}
 		cooldown, reason = upstreamCodexPlanGatedModelCooldown, upstreamCodexPlanGatedModelReason
 	default:
 		return false
@@ -2063,6 +2055,9 @@ func (s *RateLimitService) HandleUpstreamModelNotFound(ctx context.Context, acco
 	if modelKey == "" {
 		return false
 	}
+	if shouldSkipCodexPlanGatedImageModelCooldown(ctx, reason, requestedModel, modelKey) {
+		return true
+	}
 	resetAt := time.Now().Add(cooldown)
 	if err := s.accountRepo.SetModelRateLimit(ctx, account.ID, modelKey, resetAt, reason); err != nil {
 		slog.Warn("upstream_model_not_found_set_model_rate_limit_failed", "account_id", account.ID, "model", modelKey, "reason", reason, "error", err)
@@ -2072,6 +2067,29 @@ func (s *RateLimitService) HandleUpstreamModelNotFound(ctx context.Context, acco
 	return true
 }
 
+// shouldSkipCodexPlanGatedImageModelCooldown 判断这次 Codex plan-gated 400 是否
+// 属于"图片模型被文本端点拒绝"。
+//
+// 这类错误是确定性的端点错配，不是账号能力缺失：同一账号通过 /v1/images/* 依然
+// 能出图。在这里写 per-model 冷却，会让一次用错端点的请求把整个号池对正确的生图
+// 端点下线（#4828）。
+//
+// 但请求本身就从 /v1/images/* 入站时不适用——那种情况下被拒说明账号确实不具备
+// 该模型能力，冷却是必要的刹车：没有它，每个生图请求都会完整走一遍号池，对上游
+// 形成无上界的 400 放大。
+//
+// 请求模型与最终冷却键都要判：冷却键走的是 account.GetMappedModel，账号可能把
+// 文本别名映射到 gpt-image-*，只判请求模型会漏掉这种形态。
+func shouldSkipCodexPlanGatedImageModelCooldown(ctx context.Context, reason, requestedModel, modelKey string) bool {
+	if reason != upstreamCodexPlanGatedModelReason {
+		return false
+	}
+	if OpenAIImagesEndpointFromContext(ctx) {
+		return false
+	}
+	return IsGPTImageGenerationModel(requestedModel) || IsGPTImageGenerationModel(modelKey)
+}
+
 func modelRateLimitKeyForUpstreamModelNotFound(ctx context.Context, account *Account, requestedModel string) string {
 	modelKey := strings.TrimSpace(requestedModel)
 	if account == nil || modelKey == "" {
diff --git a/backend/internal/service/ratelimit_service_model_not_found_test.go b/backend/internal/service/ratelimit_service_model_not_found_test.go
index 3e22d32b60..5e2fa0f71d 100644
--- a/backend/internal/service/ratelimit_service_model_not_found_test.go
+++ b/backend/internal/service/ratelimit_service_model_not_found_test.go
@@ -435,3 +435,93 @@ func openAICodexPlanGatedOAuthAccount() *Account {
 		Credentials: map[string]any{},
 	}
 }
+
+// 请求本身就走 /v1/images/* 时必须保留冷却。
+//
+// OAuth 账号的 /v1/images/* 上游同样是 Codex Responses（openai_images_responses.go
+// → handleOpenAIImagesErrorResponse → handleOpenAIAccountUpstreamError →
+// HandleUpstreamModelNotFound），所以这条路径也会命中 plan-gated 分支。账号确实
+// 不具备生图能力时，冷却是唯一的刹车：调度层靠 model_rate_limits 跳过该账号后
+// 快速 503；一旦跳过冷却，每个请求都会完整走一遍号池，对上游形成无上界的 400 放大。
+func TestRateLimitService_HandleUpstreamError_CodexPlanGatedImageModelKeepsCooldownOnImagesEndpoint(t *testing.T) {
+	repo := &modelNotFoundAccountRepoStub{}
+	svc := &RateLimitService{accountRepo: repo}
+	account := openAICodexPlanGatedOAuthAccount()
+
+	handled := svc.HandleUpstreamError(
+		WithOpenAIImagesEndpoint(context.Background()),
+		account,
+		http.StatusBadRequest,
+		http.Header{},
+		[]byte(`{"detail":"The 'gpt-image-2' model is not supported when using Codex with a ChatGPT account."}`),
+		"gpt-image-2",
+	)
+
+	require.True(t, handled)
+	require.Len(t, repo.modelRateLimitCalls, 1,
+		"/v1/images/* 上的 plan-gated 拒绝是真实的能力缺失，必须保留冷却刹车")
+	require.Equal(t, "gpt-image-2", repo.modelRateLimitCalls[0].scope)
+	require.Equal(t, upstreamCodexPlanGatedModelReason, repo.modelRateLimitCalls[0].reason)
+}
+
+// 仅 WithOpenAIImageGenerationIntent（/v1/responses 因模型名自动置位）不算专用生图
+// 端点，仍按"用错端点"处理。
+func TestRateLimitService_HandleUpstreamError_CodexPlanGatedImageModelSkipsCooldownOnIntentOnly(t *testing.T) {
+	repo := &modelNotFoundAccountRepoStub{}
+	svc := &RateLimitService{accountRepo: repo}
+	account := openAICodexPlanGatedOAuthAccount()
+
+	handled := svc.HandleUpstreamError(
+		WithOpenAIImageGenerationIntent(context.Background()),
+		account,
+		http.StatusBadRequest,
+		http.Header{},
+		[]byte(`{"detail":"The 'gpt-image-2' model is not supported when using Codex with a ChatGPT account."}`),
+		"gpt-image-2",
+	)
+
+	require.True(t, handled)
+	require.Empty(t, repo.modelRateLimitCalls)
+}
+
+// 守卫口径必须与冷却键一致：冷却键走 account.GetMappedModel，账号可以把文本别名
+// 映射到 gpt-image-*，只判请求模型会漏掉这种形态，原 bug 原样复现。
+func TestRateLimitService_HandleUpstreamError_CodexPlanGatedImageModelSkipsCooldownViaModelMapping(t *testing.T) {
+	repo := &modelNotFoundAccountRepoStub{}
+	svc := &RateLimitService{accountRepo: repo}
+	account := openAICodexPlanGatedOAuthAccount()
+	account.Credentials["model_mapping"] = map[string]any{"my-draw-alias": "gpt-image-2"}
+
+	handled := svc.HandleUpstreamError(
+		context.Background(),
+		account,
+		http.StatusBadRequest,
+		http.Header{},
+		[]byte(`{"detail":"The 'gpt-image-2' model is not supported when using Codex with a ChatGPT account."}`),
+		"my-draw-alias",
+	)
+
+	require.True(t, handled)
+	require.Empty(t, repo.modelRateLimitCalls,
+		"映射后的上游模型是图片模型，冷却键会写到 gpt-image-2 上，守卫必须一并识别")
+}
+
+// 404 model-not-found 分支不受守卫影响：即使是图片模型也照常冷却。
+func TestRateLimitService_HandleUpstreamError_ModelNotFoundImageModelStillCoolsDown(t *testing.T) {
+	repo := &modelNotFoundAccountRepoStub{}
+	svc := &RateLimitService{accountRepo: repo}
+	account := openAICodexPlanGatedOAuthAccount()
+
+	handled := svc.HandleUpstreamError(
+		context.Background(),
+		account,
+		http.StatusNotFound,
+		http.Header{},
+		[]byte(`{"error":{"message":"The model 'gpt-image-2' does not exist","code":"model_not_found"}}`),
+		"gpt-image-2",
+	)
+
+	require.True(t, handled)
+	require.Len(t, repo.modelRateLimitCalls, 1, "守卫只作用于 codex plan-gated 分支")
+	require.Equal(t, upstreamModelNotFoundReason, repo.modelRateLimitCalls[0].reason)
+}