diff --git a/backend/internal/service/grok_upstream_errors.go b/backend/internal/service/grok_upstream_errors.go
index 880cd702b0..0039ea4cdc 100644
--- a/backend/internal/service/grok_upstream_errors.go
+++ b/backend/internal/service/grok_upstream_errors.go
@@ -191,10 +191,17 @@ func grokContentPolicyClientMessage(responseBody []byte) string {
 // shouldFailoverGrokUpstreamError is the body-aware counterpart of the
 // status-only failover helper. Grok content refusals must stay on the current
 // account and be returned to the caller instead of consuming the account pool.
+// Free-usage / empty-output / billing bodies also failover even when the HTTP
+// status alone would not (e.g. 400 with free-usage-exhausted).
 func (s *OpenAIGatewayService) shouldFailoverGrokUpstreamError(statusCode int, responseBody []byte) bool {
 	if isGrokContentPolicyRejection(statusCode, responseBody) {
 		return false
 	}
+	decision := classifyGrokUpstreamFailure(statusCode, responseBody, "")
+	switch decision.Class {
+	case GrokFailureFreeUsage, GrokFailureEmptyUpstream, GrokFailureBilling, GrokFailureModelCapacity:
+		return decision.ShouldFailover
+	}
 	return s.shouldFailoverUpstreamError(statusCode)
 }
 
diff --git a/backend/internal/service/grok_upstream_failure.go b/backend/internal/service/grok_upstream_failure.go
new file mode 100644
index 0000000000..86ed02e93b
--- /dev/null
+++ b/backend/internal/service/grok_upstream_failure.go
@@ -0,0 +1,494 @@
+package service
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"regexp"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/tidwall/gjson"
+)
+
+// Grok upstream failure classes used to decide temp-unschedulable cooldowns and
+// pre-commit account failover. Classification is body-first so free-usage and
+// empty-output wording still win when the proxy rewrites status codes.
+type GrokUpstreamFailureClass string
+
+const (
+	GrokFailureNone          GrokUpstreamFailureClass = ""
+	GrokFailureFreeUsage     GrokUpstreamFailureClass = "subscription:free-usage-exhausted"
+	GrokFailureBilling       GrokUpstreamFailureClass = "billing_quota"
+	GrokFailureEmptyUpstream GrokUpstreamFailureClass = "empty_upstream"
+	GrokFailureModelCapacity GrokUpstreamFailureClass = "model_capacity"
+	GrokFailureRateLimit     GrokUpstreamFailureClass = "rate_limit"
+	GrokFailureAuth          GrokUpstreamFailureClass = "auth_error"
+	GrokFailureServer        GrokUpstreamFailureClass = "server_error"
+)
+
+// GrokUpstreamFailureDecision is a pure classification result. Callers map it
+// onto existing account state helpers (tempUnscheduleGrok / rateLimitGrok).
+// BlockModel is retained for observability; the current scheduler does not
+// implement per-model soft-blocks, so free-usage deliberately never sets it.
+type GrokUpstreamFailureDecision struct {
+	Class          GrokUpstreamFailureClass
+	Model          string
+	Cooldown       time.Duration
+	ShouldCooldown bool
+	// ShouldFailover recommends trying another account before writing a
+	// terminal response (pre-commit only). Content-policy rejections are
+	// handled separately and never reach this classifier for failover.
+	ShouldFailover bool
+	// BlockModel is true only for empty-output when a model id is known.
+	// Free-usage never sets this: the account cools, not a single model.
+	BlockModel   bool
+	Reason       string
+	TokensActual *int64
+	TokensLimit  *int64
+}
+
+var (
+	reGrokTokenPair    = regexp.MustCompile(`(?i)tokens?\s*(?:\(actual\s*/\s*limit\))?\s*[:=]?\s*(\d+)\s*/\s*(\d+)`)
+	reGrokModelFor     = regexp.MustCompile(`(?i)(?:for\s+model|model|模型)\s*[:：]?\s*([a-z0-9][a-z0-9._-]{2,80})`)
+	reGrokResetsWindow = regexp.MustCompile(`(?i)resets?\s+over\s+a\s+rolling\s+(\d+)\s*-\s*hour`)
+)
+
+// classifyGrokUpstreamFailure decides cooldown/failover from status + body.
+// Priority (body/code first, status second):
+//  1. free-usage exhausted → account cool, no model block, failover
+//  2. billing hard quota → longer cool, failover
+//  3. empty model output → short cool + optional model soft-block marker, failover
+//  4. model capacity → short cool, failover
+//  5. bare rate-limit / 429 without free-usage language → cool, failover
+//  6. bare 5xx → brief cool, failover
+//  7. validation / client errors without quota language → no cool
+//
+// Content-policy 403s must be filtered by the caller before invoking this.
+func classifyGrokUpstreamFailure(statusCode int, responseBody []byte, requestedModel string) GrokUpstreamFailureDecision {
+	text, code, low := grokUpstreamErrorCorpus(statusCode, responseBody)
+	model := extractGrokFailureModel(text, responseBody, requestedModel)
+	actual, limit, hasTokens := parseGrokTokenPair(text)
+	if !hasTokens {
+		actual, limit, hasTokens = parseGrokTokenPair(string(responseBody))
+	}
+
+	// --- Free usage / rolling quota exhausted ---
+	if isGrokFreeUsageExhaustedText(low) || isGrokFreeUsageCode(code) || isGrokFreeUsageCode(text) {
+		d := GrokUpstreamFailureDecision{
+			Class:          GrokFailureFreeUsage,
+			Model:          model,
+			Cooldown:       grokFreeUsageCooldownDuration(low),
+			ShouldCooldown: true,
+			ShouldFailover: true,
+			BlockModel:     false,
+			Reason:         firstNonEmpty(text, code, "free usage exhausted"),
+		}
+		if hasTokens {
+			a, b := actual, limit
+			d.TokensActual = &a
+			d.TokensLimit = &b
+		}
+		return d
+	}
+
+	// Billing / hard quota (not free-tier rolling).
+	// Cooldown stays at 30m to match the existing Grok 402/spending-limit
+	// handler (longer cools would change ops behavior without a settings knob).
+	if isGrokBillingQuotaText(low) || statusCode == http.StatusPaymentRequired {
+		reason := firstNonEmpty(text, "billing quota")
+		if statusCode == http.StatusPaymentRequired && text == "" {
+			reason = "payment required"
+		}
+		return GrokUpstreamFailureDecision{
+			Class:          GrokFailureBilling,
+			Model:          model,
+			Cooldown:       30 * time.Minute,
+			ShouldCooldown: true,
+			ShouldFailover: true,
+			BlockModel:     model != "",
+			Reason:         reason,
+		}
+	}
+
+	// Empty HTTP 200 / empty model output (often rewritten to synthetic 502).
+	if isGrokEmptyModelOutputText(low) || isGrokEmptyModelOutputCode(code) {
+		return GrokUpstreamFailureDecision{
+			Class:          GrokFailureEmptyUpstream,
+			Model:          model,
+			Cooldown:       4 * time.Minute,
+			ShouldCooldown: true,
+			ShouldFailover: true,
+			BlockModel:     model != "",
+			Reason:         firstNonEmpty(text, "empty model output"),
+		}
+	}
+
+	// Model capacity / overloaded.
+	if isGrokModelCapacityText(low) {
+		return GrokUpstreamFailureDecision{
+			Class:          GrokFailureModelCapacity,
+			Model:          model,
+			Cooldown:       3 * time.Minute,
+			ShouldCooldown: true,
+			ShouldFailover: true,
+			BlockModel:     false,
+			Reason:         firstNonEmpty(text, "model capacity"),
+		}
+	}
+
+	// Rate limit without free-usage language.
+	if statusCode == http.StatusTooManyRequests || isGrokRateLimitText(low) {
+		return GrokUpstreamFailureDecision{
+			Class:          GrokFailureRateLimit,
+			Model:          model,
+			Cooldown:       10 * time.Minute,
+			ShouldCooldown: true,
+			ShouldFailover: true,
+			BlockModel:     false,
+			Reason:         firstNonEmpty(text, "rate limit"),
+		}
+	}
+
+	// Upstream 5xx — brief cool. Empty-output synthetic 502 already handled above.
+	if statusCode >= 500 && statusCode <= 599 {
+		return GrokUpstreamFailureDecision{
+			Class:          GrokFailureServer,
+			Cooldown:       2 * time.Minute,
+			ShouldCooldown: true,
+			ShouldFailover: true,
+			Reason:         firstNonEmpty(text, "server error"),
+		}
+	}
+
+	return GrokUpstreamFailureDecision{Reason: text}
+}
+
+func grokUpstreamErrorCorpus(statusCode int, responseBody []byte) (text, code, low string) {
+	raw := strings.TrimSpace(string(responseBody))
+	if unwrappedStatus, unwrappedBody, ok := unwrapGrokUpstreamErrorText(raw); ok {
+		if statusCode <= 0 {
+			statusCode = unwrappedStatus
+		}
+		raw = unwrappedBody
+	}
+	text = raw
+	codeFromJSON, msgFromJSON := parseGrokUpstreamErrorJSON(raw)
+	if msgFromJSON != "" {
+		if text == "" || len(msgFromJSON) > len(text)/2 || looksLikeGrokQuotaMessage(msgFromJSON) {
+			text = msgFromJSON
+		}
+	}
+	// Prefer structured fields from the original body when present.
+	if len(responseBody) > 0 {
+		if m := strings.TrimSpace(firstNonEmpty(
+			gjson.GetBytes(responseBody, "error.message").String(),
+			gjson.GetBytes(responseBody, "message").String(),
+			gjson.GetBytes(responseBody, "error").String(),
+		)); m != "" && (text == "" || looksLikeGrokQuotaMessage(m)) {
+			text = m
+		}
+		if c := strings.TrimSpace(firstNonEmpty(
+			gjson.GetBytes(responseBody, "error.code").String(),
+			gjson.GetBytes(responseBody, "code").String(),
+		)); c != "" {
+			codeFromJSON = c
+		}
+	}
+	code = codeFromJSON
+	low = strings.ToLower(strings.TrimSpace(text))
+	if code != "" && !strings.Contains(low, strings.ToLower(code)) {
+		low = strings.ToLower(code) + " " + low
+	}
+	return text, code, low
+}
+
+func unwrapGrokUpstreamErrorText(errText string) (status int, body string, ok bool) {
+	text := strings.TrimSpace(errText)
+	if text == "" {
+		return 0, "", false
+	}
+	lower := strings.ToLower(text)
+	for _, p := range []string{"upstream status ", "status "} {
+		if !strings.HasPrefix(lower, p) {
+			continue
+		}
+		rest := strings.TrimSpace(text[len(p):])
+		i := 0
+		for i < len(rest) && rest[i] >= '0' && rest[i] <= '9' {
+			status = status*10 + int(rest[i]-'0')
+			i++
+		}
+		if status <= 0 || i == 0 {
+			return 0, "", false
+		}
+		rest = strings.TrimSpace(rest[i:])
+		if strings.HasPrefix(rest, ":") {
+			rest = strings.TrimSpace(rest[1:])
+		}
+		return status, rest, true
+	}
+	return 0, "", false
+}
+
+func parseGrokUpstreamErrorJSON(errText string) (code, message string) {
+	text := strings.TrimSpace(errText)
+	if text == "" || text[0] != '{' {
+		return "", ""
+	}
+	var payload map[string]any
+	if json.Unmarshal([]byte(text), &payload) != nil {
+		return "", ""
+	}
+	if v, ok := payload["code"].(string); ok {
+		code = v
+	}
+	if v, ok := payload["message"].(string); ok {
+		message = v
+	}
+	if errObj, ok := payload["error"].(map[string]any); ok {
+		if v, ok := errObj["code"].(string); ok && code == "" {
+			code = v
+		}
+		if v, ok := errObj["message"].(string); ok && message == "" {
+			message = v
+		}
+	}
+	if errStr, ok := payload["error"].(string); ok && message == "" {
+		message = errStr
+	}
+	return strings.TrimSpace(code), strings.TrimSpace(message)
+}
+
+func looksLikeGrokQuotaMessage(s string) bool {
+	low := strings.ToLower(s)
+	return strings.Contains(low, "quota") ||
+		strings.Contains(low, "usage") ||
+		strings.Contains(low, "credit") ||
+		strings.Contains(low, "额度") ||
+		strings.Contains(low, "free")
+}
+
+func isGrokFreeUsageCode(code string) bool {
+	c := strings.ToLower(strings.TrimSpace(code))
+	if c == "" {
+		return false
+	}
+	if strings.Contains(c, "subscription:free-usage-exhausted") ||
+		strings.Contains(c, "free-usage-exhausted") ||
+		strings.Contains(c, "free_usage_exhausted") ||
+		strings.Contains(c, "usage-limit-exceeded") ||
+		strings.Contains(c, "usage_limit_exceeded") {
+		return true
+	}
+	return (strings.Contains(c, "free-usage") || strings.Contains(c, "free_usage")) &&
+		(strings.Contains(c, "exhaust") || strings.Contains(c, "exceed") || strings.Contains(c, "limit"))
+}
+
+func isGrokFreeUsageExhaustedText(low string) bool {
+	if low == "" {
+		return false
+	}
+	if strings.Contains(low, "free-usage-exhausted") ||
+		strings.Contains(low, "free_usage_exhausted") ||
+		strings.Contains(low, "subscription:free-usage") ||
+		strings.Contains(low, "usage-limit-exceeded") ||
+		strings.Contains(low, "usage_limit_exceeded") ||
+		strings.Contains(low, "free-tier-limit") ||
+		strings.Contains(low, "free_tier_limit") {
+		return true
+	}
+	if strings.Contains(low, "free usage") ||
+		strings.Contains(low, "included free usage") ||
+		strings.Contains(low, "used all the included free") ||
+		strings.Contains(low, "you've used all the included free") ||
+		strings.Contains(low, "you have used all the included free") ||
+		strings.Contains(low, "free quota") ||
+		strings.Contains(low, "no remaining free") ||
+		strings.Contains(low, "out of free") ||
+		strings.Contains(low, "usage resets over a rolling") ||
+		(strings.Contains(low, "free tier") && (strings.Contains(low, "exhaust") || strings.Contains(low, "limit") || strings.Contains(low, "exceed"))) {
+		return true
+	}
+	for _, p := range []string{
+		"额度耗尽", "额度用完", "额度不足", "额度已用尽", "额度已耗尽",
+		"免费额度", "免费用量", "用量用完", "用量耗尽", "用量超限", "用量已用尽",
+		"配额耗尽", "配额已用尽", "配额不足", "配额超限", "配额用完",
+		"没有额度", "没额度", "无额度", "可用额度不足", "模型额度",
+		"临时额度", "额度已满", "额度超限", "额度达到上限",
+		"模型额度用完", "模型额度耗尽", "账号额度用完", "账号额度耗尽",
+		"额度不够", "没额度了", "额度没了", "用完额度", "耗尽额度",
+	} {
+		if strings.Contains(low, p) {
+			return true
+		}
+	}
+	if (strings.Contains(low, "quota") && (strings.Contains(low, "exhaust") || strings.Contains(low, "exceed") || strings.Contains(low, "limit"))) ||
+		(strings.Contains(low, "usage") && (strings.Contains(low, "exhaust") || strings.Contains(low, "exceed")) && (strings.Contains(low, "limit") || strings.Contains(low, "free") || strings.Contains(low, "model"))) {
+		if strings.Contains(low, "free") || strings.Contains(low, "rolling") ||
+			strings.Contains(low, "24-hour") || strings.Contains(low, "24 hour") ||
+			strings.Contains(low, "model") || strings.Contains(low, "subscription") ||
+			strings.Contains(low, "included") || strings.Contains(low, "tokens") {
+			return true
+		}
+	}
+	if a, b, ok := parseGrokTokenPair(low); ok && b > 0 && a >= b {
+		if strings.Contains(low, "free") || strings.Contains(low, "subscription") ||
+			strings.Contains(low, "included") || strings.Contains(low, "model") ||
+			strings.Contains(low, "usage") || strings.Contains(low, "quota") ||
+			strings.Contains(low, "rolling") {
+			return true
+		}
+	}
+	return false
+}
+
+func isGrokBillingQuotaText(low string) bool {
+	if low == "" {
+		return false
+	}
+	if strings.Contains(low, "insufficient_quota") {
+		return true
+	}
+	if strings.Contains(low, "billing") && strings.Contains(low, "quota") {
+		return true
+	}
+	if strings.Contains(low, "payment") && (strings.Contains(low, "required") || strings.Contains(low, "fail")) {
+		return true
+	}
+	if strings.Contains(low, "spending limit") || strings.Contains(low, "run out of credits") || strings.Contains(low, "out of credits") {
+		return true
+	}
+	if strings.Contains(low, "余额不足") || strings.Contains(low, "欠费") || strings.Contains(low, "需要付费") {
+		return true
+	}
+	return false
+}
+
+func isGrokModelCapacityText(low string) bool {
+	return strings.Contains(low, "capacity") ||
+		strings.Contains(low, "overloaded") ||
+		strings.Contains(low, "server_busy") ||
+		strings.Contains(low, "too many concurrent") ||
+		strings.Contains(low, "engine_overloaded")
+}
+
+func isGrokRateLimitText(low string) bool {
+	return strings.Contains(low, "rate limit") ||
+		strings.Contains(low, "rate_limit") ||
+		strings.Contains(low, "too many requests") ||
+		strings.Contains(low, "请求过于频繁") ||
+		strings.Contains(low, "速率限制")
+}
+
+func isGrokEmptyModelOutputText(low string) bool {
+	if low == "" {
+		return false
+	}
+	return strings.Contains(low, "empty model output") ||
+		strings.Contains(low, "no content/tool_calls") ||
+		strings.Contains(low, "no client-visible content") ||
+		strings.Contains(low, "empty_upstream") ||
+		strings.Contains(low, "empty upstream")
+}
+
+func isGrokEmptyModelOutputCode(code string) bool {
+	c := strings.ToLower(strings.TrimSpace(code))
+	if c == "" {
+		return false
+	}
+	return c == "empty_upstream" ||
+		c == "empty-model-output" ||
+		c == "empty_model_output" ||
+		strings.Contains(c, "empty_upstream") ||
+		strings.Contains(c, "empty-model-output")
+}
+
+func grokFreeUsageCooldownDuration(low string) time.Duration {
+	if m := reGrokResetsWindow.FindStringSubmatch(low); len(m) == 2 {
+		hours, _ := strconv.Atoi(m[1])
+		if hours > 0 {
+			d := time.Duration(hours) * time.Hour / 12
+			if d < 20*time.Minute {
+				d = 20 * time.Minute
+			}
+			if d > 6*time.Hour {
+				d = 6 * time.Hour
+			}
+			return d
+		}
+	}
+	return 2 * time.Hour
+}
+
+func parseGrokTokenPair(errText string) (actual, limit int64, ok bool) {
+	m := reGrokTokenPair.FindStringSubmatch(errText)
+	if len(m) != 3 {
+		return 0, 0, false
+	}
+	a, errA := strconv.ParseInt(m[1], 10, 64)
+	b, errB := strconv.ParseInt(m[2], 10, 64)
+	if errA != nil || errB != nil {
+		return 0, 0, false
+	}
+	return a, b, true
+}
+
+func extractGrokFailureModel(text string, responseBody []byte, fallback string) string {
+	if m := reGrokModelFor.FindStringSubmatch(text); len(m) == 2 {
+		return strings.TrimSpace(m[1])
+	}
+	if len(responseBody) > 0 {
+		if m := strings.TrimSpace(firstNonEmpty(
+			gjson.GetBytes(responseBody, "error.model").String(),
+			gjson.GetBytes(responseBody, "model").String(),
+		)); m != "" {
+			return m
+		}
+	}
+	return strings.TrimSpace(fallback)
+}
+
+// applyGrokUpstreamFailureDecision maps a classification onto existing account
+// health helpers. Returns true when the decision fully handled the error path
+// (caller should not apply the status-code switch defaults again).
+func (s *OpenAIGatewayService) applyGrokUpstreamFailureDecision(
+	ctx context.Context,
+	account *Account,
+	decision GrokUpstreamFailureDecision,
+) bool {
+	if s == nil || account == nil || !decision.ShouldCooldown || decision.Cooldown <= 0 {
+		return false
+	}
+	// Keep reasons short and stable for ops UI / temp_unschedulable_reason.
+	var reason string
+	switch decision.Class {
+	case GrokFailureFreeUsage:
+		reason = "grok free usage exhausted"
+	case GrokFailureBilling:
+		low := strings.ToLower(decision.Reason)
+		if strings.Contains(low, "spending") || strings.Contains(low, "credits") {
+			reason = "grok spending limit"
+		} else {
+			// Keep the historical 402/payment reason for ops UI + regression tests.
+			reason = "grok payment required"
+		}
+	case GrokFailureEmptyUpstream:
+		reason = "grok empty model output"
+	case GrokFailureModelCapacity:
+		reason = "grok model capacity"
+	case GrokFailureRateLimit:
+		// Pure 429 without free-usage language keeps the existing rate-limit
+		// snapshot path (Retry-After / quota headers). Body-only rate-limit
+		// phrasing still cools here via ShouldCooldown from the classifier, but
+		// the handler only invokes this for non-RateLimit classes.
+		return false
+	case GrokFailureServer:
+		reason = "grok upstream temporary error"
+	default:
+		return false
+	}
+	s.tempUnscheduleGrok(ctx, account, decision.Cooldown, reason)
+	return true
+}
diff --git a/backend/internal/service/grok_upstream_failure_test.go b/backend/internal/service/grok_upstream_failure_test.go
new file mode 100644
index 0000000000..c51b4053a6
--- /dev/null
+++ b/backend/internal/service/grok_upstream_failure_test.go
@@ -0,0 +1,167 @@
+//go:build unit
+
+package service
+
+import (
+	"context"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestClassifyGrokUpstreamFailure_FreeUsage(t *testing.T) {
+	cases := []struct {
+		name   string
+		status int
+		body   string
+	}{
+		{
+			name:   "code free-usage-exhausted",
+			status: http.StatusTooManyRequests,
+			body:   `{"error":{"code":"subscription:free-usage-exhausted","message":"You've used all the included free usage for model grok-4.5. Usage resets over a rolling 24-hour window."}}`,
+		},
+		{
+			name:   "chinese body without 429",
+			status: http.StatusBadRequest,
+			body:   `{"error":{"message":"模型额度用完，请稍后再试"}}`,
+		},
+		{
+			name:   "token pair with free marker",
+			status: http.StatusOK,
+			body:   `{"error":{"message":"free usage tokens (actual / limit): 2000000 / 2000000 for model grok-4.5"}}`,
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			d := classifyGrokUpstreamFailure(tc.status, []byte(tc.body), "grok-4.5")
+			require.Equal(t, GrokFailureFreeUsage, d.Class)
+			require.True(t, d.ShouldCooldown)
+			require.True(t, d.ShouldFailover)
+			require.False(t, d.BlockModel, "free-usage must not soft-block models")
+			require.GreaterOrEqual(t, d.Cooldown, 20*time.Minute)
+		})
+	}
+}
+
+func TestClassifyGrokUpstreamFailure_EmptyUpstream(t *testing.T) {
+	d := classifyGrokUpstreamFailure(http.StatusBadGateway, []byte(`empty model output: no content/tool_calls`), "grok-4.5")
+	require.Equal(t, GrokFailureEmptyUpstream, d.Class)
+	require.True(t, d.ShouldCooldown)
+	require.True(t, d.ShouldFailover)
+	require.True(t, d.BlockModel)
+	require.Equal(t, 4*time.Minute, d.Cooldown)
+}
+
+func TestClassifyGrokUpstreamFailure_Billing(t *testing.T) {
+	d := classifyGrokUpstreamFailure(http.StatusForbidden, []byte(`{"code":"personal-team-blocked:spending-limit","error":"spending limit reached"}`), "")
+	require.Equal(t, GrokFailureBilling, d.Class)
+	require.True(t, d.ShouldCooldown)
+	require.True(t, d.ShouldFailover)
+}
+
+func TestClassifyGrokUpstreamFailure_ValidationNoCool(t *testing.T) {
+	d := classifyGrokUpstreamFailure(http.StatusBadRequest, []byte(`{"error":{"message":"invalid tool schema"}}`), "")
+	require.Equal(t, GrokFailureNone, d.Class)
+	require.False(t, d.ShouldCooldown)
+	require.False(t, d.ShouldFailover)
+}
+
+func TestClassifyGrokUpstreamFailure_FreeUsageWinsOver5xx(t *testing.T) {
+	// Proxy may rewrite free-usage into synthetic 502; body must win.
+	d := classifyGrokUpstreamFailure(http.StatusBadGateway, []byte(`subscription:free-usage-exhausted for model grok-4.3`), "grok-4.3")
+	require.Equal(t, GrokFailureFreeUsage, d.Class)
+	require.NotEqual(t, GrokFailureServer, d.Class)
+}
+
+func TestShouldFailoverGrokUpstreamError_FreeUsageBody(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	body := []byte(`{"error":{"code":"subscription:free-usage-exhausted","message":"free usage exhausted"}}`)
+	require.True(t, svc.shouldFailoverGrokUpstreamError(http.StatusBadRequest, body))
+}
+
+func TestShouldFailoverGrokUpstreamError_ContentPolicyStillNoFailover(t *testing.T) {
+	svc := &OpenAIGatewayService{}
+	body := []byte(`{"error":{"code":"new_sensitive","message":"text is sensitive"}}`)
+	require.False(t, svc.shouldFailoverGrokUpstreamError(http.StatusForbidden, body))
+}
+
+func TestHandleGrokAccountUpstreamError_FreeUsageBodyCoolsAccount(t *testing.T) {
+	repo := &grokQuotaAccountRepo{}
+	svc := &OpenAIGatewayService{accountRepo: repo}
+	account := &Account{ID: 9101, Platform: PlatformGrok, Type: AccountTypeOAuth}
+	before := time.Now()
+	body := []byte(`{"error":{"code":"subscription:free-usage-exhausted","message":"You've used all the included free usage. Usage resets over a rolling 24-hour window."}}`)
+
+	svc.handleGrokAccountUpstreamError(context.Background(), account, http.StatusBadRequest, nil, body)
+
+	require.Equal(t, 1, repo.tempUnschedCalls)
+	require.Equal(t, "grok free usage exhausted", repo.lastTempUnschedReason)
+	// 24h rolling → 2h cool
+	require.Greater(t, repo.lastTempUnschedUntil, before.Add(119*time.Minute))
+	require.Less(t, repo.lastTempUnschedUntil, before.Add(121*time.Minute))
+}
+
+func TestHandleGrokAccountUpstreamError_EmptyOutputCoolsAccount(t *testing.T) {
+	repo := &grokQuotaAccountRepo{}
+	svc := &OpenAIGatewayService{accountRepo: repo}
+	account := &Account{ID: 9102, Platform: PlatformGrok, Type: AccountTypeOAuth}
+	before := time.Now()
+
+	svc.handleGrokAccountUpstreamError(
+		context.Background(), account, http.StatusBadGateway, nil,
+		[]byte(`empty model output: no content/tool_calls`),
+	)
+
+	require.Equal(t, 1, repo.tempUnschedCalls)
+	require.Equal(t, "grok empty model output", repo.lastTempUnschedReason)
+	require.WithinDuration(t, before.Add(4*time.Minute), repo.lastTempUnschedUntil, time.Second)
+}
+
+func TestHandleGrokAccountUpstreamError_FreeUsageDoesNotCoolPoolMode(t *testing.T) {
+	repo := &grokQuotaAccountRepo{}
+	svc := &OpenAIGatewayService{accountRepo: repo}
+	account := &Account{
+		ID:       9103,
+		Platform: PlatformGrok,
+		Type:     AccountTypeAPIKey,
+		Credentials: map[string]any{
+			"pool_mode": true,
+		},
+	}
+	body := []byte(`{"error":{"code":"subscription:free-usage-exhausted","message":"free usage exhausted"}}`)
+
+	svc.handleGrokAccountUpstreamError(context.Background(), account, http.StatusBadRequest, nil, body)
+
+	require.Zero(t, repo.tempUnschedCalls)
+	require.False(t, svc.isOpenAIAccountRuntimeBlocked(account))
+}
+
+func TestHandleGrokAccountUpstreamError_ContentPolicyStillNoMutation(t *testing.T) {
+	repo := &grokQuotaAccountRepo{}
+	svc := &OpenAIGatewayService{accountRepo: repo}
+	account := &Account{ID: 9104, Platform: PlatformGrok, Type: AccountTypeOAuth}
+	body := []byte(`{"error":{"code":"new_sensitive","message":"text is sensitive"}}`)
+
+	svc.handleGrokAccountUpstreamError(context.Background(), account, http.StatusForbidden, nil, body)
+
+	require.Zero(t, repo.tempUnschedCalls)
+}
+
+func TestHandleGrokAccountUpstreamError_Entitlement403Unchanged(t *testing.T) {
+	repo := &grokQuotaAccountRepo{}
+	svc := &OpenAIGatewayService{accountRepo: repo}
+	account := &Account{ID: 9105, Platform: PlatformGrok, Type: AccountTypeOAuth}
+	before := time.Now()
+
+	svc.handleGrokAccountUpstreamError(
+		context.Background(), account, http.StatusForbidden, nil,
+		[]byte(`{"error":{"message":"subscription required"}}`),
+	)
+
+	require.Equal(t, 1, repo.tempUnschedCalls)
+	require.Equal(t, "grok access or entitlement denied", repo.lastTempUnschedReason)
+	require.Greater(t, repo.lastTempUnschedUntil, before.Add(29*time.Minute))
+	require.Less(t, repo.lastTempUnschedUntil, before.Add(31*time.Minute))
+}
diff --git a/backend/internal/service/openai_gateway_grok.go b/backend/internal/service/openai_gateway_grok.go
index 0d3f024628..8140c14cf8 100644
--- a/backend/internal/service/openai_gateway_grok.go
+++ b/backend/internal/service/openai_gateway_grok.go
@@ -1356,6 +1356,19 @@ func (s *OpenAIGatewayService) handleGrokAccountUpstreamError(ctx context.Contex
 	}
 	now := time.Now()
 	s.updateGrokUsageSnapshot(ctx, account, parseGrokQuotaSnapshot(headers, statusCode, now))
+
+	// Body-first free-usage / empty / billing / capacity must run before the
+	// status switch so non-429 free-usage bodies still cool the account.
+	// Pool-mode still skips durable mutation unless an explicit temp rule matches.
+	decision := classifyGrokUpstreamFailure(statusCode, responseBody, "")
+	if decision.ShouldCooldown && decision.Class != GrokFailureNone && decision.Class != GrokFailureRateLimit {
+		if account.IsPoolMode() {
+			// Allow configured temp rules (403) below; skip default body cools.
+		} else if s.applyGrokUpstreamFailureDecision(ctx, account, decision) {
+			return
+		}
+	}
+
 	if statusCode == http.StatusForbidden && s.applyGrokForbiddenPolicy(ctx, account, responseBody) {
 		return
 	}
@@ -1367,11 +1380,10 @@ func (s *OpenAIGatewayService) handleGrokAccountUpstreamError(ctx context.Contex
 	case http.StatusUnauthorized:
 		s.tempUnscheduleGrok(ctx, account, 10*time.Minute, "grok credentials unauthorized")
 	case http.StatusPaymentRequired:
-		// 402: temporarily unschedulable with a clear payment-required reason.
+		// 402 without a body-classified billing decision: keep the legacy 30m cool.
 		s.tempUnscheduleGrok(ctx, account, 30*time.Minute, "grok payment required")
 	case http.StatusForbidden:
-		// Spending-limit 403 (personal-team-blocked:spending-limit) is billing exhaustion,
-		// not a generic entitlement denial — still temp-unschedule with a distinct reason.
+		// Spending-limit already handled by body classifier when phrasing matches.
 		if isGrokSpendingLimitError(responseBody) {
 			s.tempUnscheduleGrok(ctx, account, 30*time.Minute, "grok spending limit")
 			return
@@ -1379,6 +1391,7 @@ func (s *OpenAIGatewayService) handleGrokAccountUpstreamError(ctx context.Contex
 		s.tempUnscheduleGrok(ctx, account, 30*time.Minute, "grok access or entitlement denied")
 	case http.StatusTooManyRequests:
 		// updateGrokUsageSnapshot installs rate-limit state for non-pool accounts.
+		// Free-usage 429 was already cooled above via body classification.
 	default:
 		if statusCode >= 500 {
 			s.tempUnscheduleGrok(ctx, account, 2*time.Minute, "grok upstream temporary error")
diff --git a/backend/internal/service/openai_gateway_grok_cache.go b/backend/internal/service/openai_gateway_grok_cache.go
index 4c6923e799..696e7352ec 100644
--- a/backend/internal/service/openai_gateway_grok_cache.go
+++ b/backend/internal/service/openai_gateway_grok_cache.go
@@ -118,6 +118,12 @@ func explicitGrokCacheSeed(c *gin.Context, body []byte, explicitKey string) stri
 	if seed == "" {
 		seed = strings.TrimSpace(explicitKey)
 	}
+	// previous_response_id is last-resort: multi-turn Responses without an
+	// explicit session still share one cache identity (model is already in the
+	// isolated seed). Message ids are rejected by the seed helper.
+	if seed == "" && len(body) > 0 {
+		seed = grokPreviousResponseSessionSeed(body)
+	}
 	return seed
 }
 
diff --git a/backend/internal/service/openai_gateway_grok_cache_test.go b/backend/internal/service/openai_gateway_grok_cache_test.go
index dc3beaa461..3251bb7485 100644
--- a/backend/internal/service/openai_gateway_grok_cache_test.go
+++ b/backend/internal/service/openai_gateway_grok_cache_test.go
@@ -24,6 +24,36 @@ func newGrokCacheTestContext(apiKeyID int64) *gin.Context {
 	return c
 }
 
+func TestGrokPreviousResponseSessionSeed(t *testing.T) {
+	require.Equal(t, "grok-prev-resp:resp_abc123", grokPreviousResponseSessionSeed([]byte(`{"previous_response_id":"resp_abc123"}`)))
+	require.Empty(t, grokPreviousResponseSessionSeed([]byte(`{"previous_response_id":"msg_abc123"}`)))
+	require.Empty(t, grokPreviousResponseSessionSeed([]byte(`{"previous_response_id":""}`)))
+	require.Empty(t, grokPreviousResponseSessionSeed([]byte(`{}`)))
+}
+
+func TestResolveGrokCacheIdentityUsesPreviousResponseIDWhenNoOtherSeed(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	c := newGrokCacheTestContext(301)
+	// No prompt_cache_key / headers / reusable prefix — only previous_response_id.
+	body := []byte(`{"model":"grok","input":[{"role":"user","content":"follow up"}],"previous_response_id":"resp_chain_001"}`)
+	got := resolveGrokCacheIdentity(c, body, "", "grok-4.5")
+	require.NotEmpty(t, got)
+
+	// Same previous_response_id → same identity (model already in isolated seed).
+	again := resolveGrokCacheIdentity(c, body, "", "grok-4.5")
+	require.Equal(t, got, again)
+
+	// Different model → different identity (model scope).
+	otherModel := resolveGrokCacheIdentity(c, body, "", "grok-4.3")
+	require.NotEqual(t, got, otherModel)
+
+	// prompt_cache_key still wins over previous_response_id.
+	withCache := []byte(`{"model":"grok","prompt_cache_key":"client-session","previous_response_id":"resp_chain_001","input":[{"role":"user","content":"x"}]}`)
+	cacheID := resolveGrokCacheIdentity(c, withCache, "", "grok-4.5")
+	require.NotEmpty(t, cacheID)
+	require.NotEqual(t, got, cacheID)
+}
+
 func TestResolveGrokCacheIdentityStableAcrossAppendOnlyTurns(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 	c := newGrokCacheTestContext(101)
diff --git a/backend/internal/service/openai_gateway_scheduling.go b/backend/internal/service/openai_gateway_scheduling.go
index 23c04daf0a..0baf9f2e8e 100644
--- a/backend/internal/service/openai_gateway_scheduling.go
+++ b/backend/internal/service/openai_gateway_scheduling.go
@@ -75,6 +75,12 @@ func explicitOpenAISessionID(c *gin.Context, body []byte) string {
 // with Grok's native conversation header only for requests authenticated to a
 // Grok group. This keeps an unrelated x-grok-conv-id header from changing
 // scheduling or upstream session behavior for non-Grok groups.
+//
+// For Grok groups only, previous_response_id is a last-resort sticky seed so
+// multi-turn Responses chains stay on the same OAuth account when no explicit
+// session/conversation/prompt_cache_key is present. Non-Grok groups omit this
+// so HTTP OpenAI paths that delete previous_response_id before upstream are
+// unchanged.
 func explicitOpenAIRequestSessionID(c *gin.Context, body []byte) string {
 	if c == nil {
 		return ""
@@ -87,9 +93,27 @@ func explicitOpenAIRequestSessionID(c *gin.Context, body []byte) string {
 	if sessionID == "" && len(body) > 0 {
 		sessionID = strings.TrimSpace(gjson.GetBytes(body, "prompt_cache_key").String())
 	}
+	if sessionID == "" && isGrokRequestContext(c) && len(body) > 0 {
+		sessionID = grokPreviousResponseSessionSeed(body)
+	}
 	return sessionID
 }
 
+// grokPreviousResponseSessionSeed returns a stable sticky seed from a Responses
+// previous_response_id. Only resp_* response ids are accepted; message ids and
+// unknown shapes must not pin sticky routing or prompt-cache identity.
+func grokPreviousResponseSessionSeed(body []byte) string {
+	id := strings.TrimSpace(gjson.GetBytes(body, "previous_response_id").String())
+	if id == "" {
+		return ""
+	}
+	if ClassifyOpenAIPreviousResponseIDKind(id) != OpenAIPreviousResponseIDKindResponseID {
+		return ""
+	}
+	// Namespace so content-derived seeds never collide with response ids.
+	return "grok-prev-resp:" + id
+}
+
 // GenerateExplicitSessionHash generates a sticky-session hash only from explicit
 // client session signals. It intentionally skips content-derived fallback and is
 // used by stateless endpoints such as /v1/images.
diff --git a/frontend/src/components/account/OAuthAuthorizationFlow.vue b/frontend/src/components/account/OAuthAuthorizationFlow.vue
index 98ae7fe519..129f04bd0f 100644
--- a/frontend/src/components/account/OAuthAuthorizationFlow.vue
+++ b/frontend/src/components/account/OAuthAuthorizationFlow.vue
@@ -924,6 +924,11 @@ interface Props {
   showEmailPasswordOption?: boolean
   showManualOption?: boolean
   initialInputMethod?: AuthInputMethod
+  /**
+   * Prefill for Grok email----password reauth. Password is never stored;
+   * pass only the email (or "email----") so the operator types the password.
+   */
+  initialEmailPassword?: string
   platform?: AccountPlatform // Platform type for different UI/text
   showProjectId?: boolean // New prop to control project ID visibility
 }
@@ -949,6 +954,7 @@ const props = withDefaults(defineProps<Props>(), {
   showEmailPasswordOption: false,
   showManualOption: true,
   initialInputMethod: 'manual',
+  initialEmailPassword: '',
   platform: 'anthropic',
   showProjectId: true
 })
@@ -1010,7 +1016,7 @@ const sessionTokenInput = ref('')
 const codexSessionInput = ref('')
 const codexPATInput = ref('')
 const ssoCookieInput = ref('')
-const emailPasswordInput = ref('')
+const emailPasswordInput = ref(props.initialEmailPassword || '')
 const showHelpDialog = ref(false)
 const oauthState = ref('')
 const projectId = ref('')
@@ -1085,6 +1091,16 @@ watch(() => props.initialInputMethod, (newVal) => {
   inputMethod.value = newVal
 })
 
+watch(
+  () => props.initialEmailPassword,
+  (newVal) => {
+    // Only prefill when the field is empty so we never overwrite operator input.
+    if (newVal && !emailPasswordInput.value.trim()) {
+      emailPasswordInput.value = newVal
+    }
+  }
+)
+
 watch(inputMethod, (newVal) => {
   emit('update:inputMethod', newVal)
 })
diff --git a/frontend/src/components/admin/account/ReAuthAccountModal.vue b/frontend/src/components/admin/account/ReAuthAccountModal.vue
index 028c229d6a..a476a5684c 100644
--- a/frontend/src/components/admin/account/ReAuthAccountModal.vue
+++ b/frontend/src/components/admin/account/ReAuthAccountModal.vue
@@ -137,6 +137,8 @@
         :method-label="t('admin.accounts.inputMethod')"
         :platform="isOpenAI ? 'openai' : isGemini ? 'gemini' : isAntigravity ? 'antigravity' : isGrok ? 'grok' : 'anthropic'"
         :show-project-id="isGemini && geminiOAuthType === 'code_assist'"
+        :initial-input-method="grokInitialInputMethod"
+        :initial-email-password="grokPrefillEmailPassword"
         @generate-url="handleGenerateUrl"
         @cookie-auth="handleCookieAuth"
         @validate-refresh-token="handleGrokValidateRefreshToken"
@@ -255,6 +257,40 @@ const isAnthropic = computed(() => props.account?.platform === 'anthropic')
 const isAntigravity = computed(() => props.account?.platform === 'antigravity')
 const isGrok = computed(() => props.account?.platform === 'grok')
 
+/** Stored Grok email for reauth prefill (password is never stored). */
+const grokAccountEmail = computed(() => {
+  if (!isGrok.value || !props.account) return ''
+  const creds = (props.account.credentials || {}) as Record<string, unknown>
+  const email = typeof creds.email === 'string' ? creds.email.trim() : ''
+  return email
+})
+
+/**
+ * Prefill "email----" so the operator only types the password.
+ * Empty when no email is known (full email----password required).
+ */
+const grokPrefillEmailPassword = computed(() => {
+  const email = grokAccountEmail.value
+  return email ? `${email}----` : ''
+})
+
+/**
+ * Grok reauth default tab:
+ * - password first when we know the email (common reauth path)
+ * - refresh_token when email unknown but RT may still work
+ * - email_password otherwise
+ */
+const grokInitialInputMethod = computed<AuthInputMethod>(() => {
+  if (!isGrok.value) return 'manual'
+  if (grokAccountEmail.value) return 'email_password'
+  const creds = (props.account?.credentials || {}) as Record<string, unknown>
+  const hasRT =
+    (typeof creds.refresh_token === 'string' && creds.refresh_token.trim() !== '') ||
+    (typeof creds.has_refresh_token === 'boolean' && creds.has_refresh_token)
+  if (hasRT) return 'refresh_token'
+  return 'email_password'
+})
+
 // Computed - current OAuth state based on platform
 const currentAuthUrl = computed(() => {
   if (isOpenAILike.value) return openaiOAuth.authUrl.value
diff --git a/frontend/src/components/admin/account/__tests__/ReAuthAccountModal.grok.spec.ts b/frontend/src/components/admin/account/__tests__/ReAuthAccountModal.grok.spec.ts
index dd8c368766..1291354c59 100644
--- a/frontend/src/components/admin/account/__tests__/ReAuthAccountModal.grok.spec.ts
+++ b/frontend/src/components/admin/account/__tests__/ReAuthAccountModal.grok.spec.ts
@@ -32,4 +32,13 @@ describe('ReAuthAccountModal Grok re-auth paths', () => {
     expect(source).toContain("method === 'email_password'")
     expect(source).toContain("method === 'refresh_token'")
   })
+
+  it('prefills email---- and defaults to password method when email is known', () => {
+    expect(source).toContain('grokPrefillEmailPassword')
+    expect(source).toContain('grokInitialInputMethod')
+    expect(source).toContain(':initial-email-password="grokPrefillEmailPassword"')
+    expect(source).toContain(':initial-input-method="grokInitialInputMethod"')
+    expect(source).toContain('email----')
+    expect(source).toContain("return 'email_password'")
+  })
 })