diff --git a/internal/agent/agents.go b/internal/agent/agents.go
index 41a1fc75..edbf3f4e 100644
--- a/internal/agent/agents.go
+++ b/internal/agent/agents.go
@@ -311,7 +311,7 @@ func piModelJSON(m catalog.Model, gw string, native bool) map[string]any {
 	if m.Images {
 		e["input"] = []string{"text", "image"}
 	}
-	if levels := piThinkingLevels(m.Efforts, e["api"] == "anthropic-messages"); levels != nil {
+	if levels := piThinkingLevels(m.Efforts, e["api"] == "anthropic-messages" && gateway.ThinksOnlyWhenAsked(m.ID)); levels != nil {
 		e["thinkingLevelMap"] = levels
 	}
 	// without it Pi takes every model for a 128K one, and compacts
@@ -354,11 +354,14 @@ func openCodeVariants(efforts []string) map[string]any {
 // turned away (#243). Pi's off is the model's none when it takes one. A
 // model without none has off hidden — Pi's off asks a Responses model for
 // effort none and leaves a Chat model on the vendor's default thinking —
-// except on Anthropic's Messages API, where Pi's off sends thinking
-// disabled, which Claude takes whatever its levels; there off is left out,
-// for Pi to offer as before. A model whose levels magpie doesn't know gets
-// no map: Pi's own defaults, as before.
-func piThinkingLevels(efforts []string, anthropic bool) map[string]any {
+// except a Claude on Anthropic's Messages API (claudeOff), where Pi's off
+// sends thinking disabled, which Claude takes whatever its levels; there
+// off is left out, for Pi to offer as before. Another vendor's model there
+// keeps off hidden: without none it always thinks, and ZCode's
+// GLM-5.3-Flash turned Pi's off away ("该模型始终支持思考，不可关闭", #699),
+// as a level left out is one Pi offers. A model whose levels magpie doesn't
+// know gets no map: Pi's own defaults, as before.
+func piThinkingLevels(efforts []string, claudeOff bool) map[string]any {
 	if len(efforts) == 0 {
 		return nil
 	}
@@ -374,7 +377,7 @@ func piThinkingLevels(efforts []string, anthropic bool) map[string]any {
 			levels[l] = nil
 		}
 	}
-	if levels["off"] == nil && anthropic {
+	if levels["off"] == nil && claudeOff {
 		delete(levels, "off")
 	}
 	return levels
diff --git a/internal/agent/pi_test.go b/internal/agent/pi_test.go
index 2c6e5ac8..a185c81a 100644
--- a/internal/agent/pi_test.go
+++ b/internal/agent/pi_test.go
@@ -3,6 +3,8 @@ package agent
 import (
 	"reflect"
 	"testing"
+
+	"github.com/yetone/magpie/internal/catalog"
 )
 
 // Every one of Pi's levels is in the map, the model's own mapped and the
@@ -36,3 +38,39 @@ func TestPiThinkingLevels(t *testing.T) {
 		}
 	}
 }
+
+// #699: ZCode's GLM-5.3-Flash, on Anthropic's Messages API, always thinks
+// (its levels low, high and max, no none) and turns thinking disabled away
+// ("该模型始终支持思考，不可关闭"); its entry left off out, which Pi offers,
+// so off is null for it. A Claude there still has off left for Pi to offer,
+// and a vendor's model that takes none maps off to it.
+func TestPiOffOnMessagesAPI(t *testing.T) {
+	for _, c := range []struct {
+		id      string
+		efforts []string
+		off     any // the map's off; "left out" when it has none
+		offered []string
+	}{
+		{"zcode/GLM-5.3-Flash", []string{"low", "high", "max"}, nil, []string{"low", "high", "max"}},
+		{"zcode/GLM-5.2", []string{"none", "high", "max"}, "none", []string{"off", "high", "max"}},
+		{"anth/claude-sonnet-5", []string{"low", "medium", "high", "xhigh", "max"}, "left out", []string{"off", "low", "medium", "high", "xhigh", "max"}},
+		{"relay/opus-5.5", []string{"low", "high", "max"}, "left out", []string{"off", "low", "high", "max"}},
+	} {
+		m := catalog.Model{ID: c.id, Name: c.id, Efforts: c.efforts, APIs: []string{"anthropic"}}
+		e := piModelJSON(m, "http://127.0.0.1:1", true)
+		if e["api"] != "anthropic-messages" {
+			t.Fatalf("%s on %v", c.id, e["api"])
+		}
+		levels := e["thinkingLevelMap"].(map[string]any)
+		off, set := levels["off"]
+		if !set {
+			off = "left out"
+		}
+		if off != c.off {
+			t.Errorf("%s: off is %v, want %v", c.id, off, c.off)
+		}
+		if got := piSupported(e); !reflect.DeepEqual(got, c.offered) {
+			t.Errorf("%s: Pi offers %v, want %v", c.id, got, c.offered)
+		}
+	}
+}
diff --git a/internal/gateway/anthropic.go b/internal/gateway/anthropic.go
index c5282abc..1d6f7db2 100644
--- a/internal/gateway/anthropic.go
+++ b/internal/gateway/anthropic.go
@@ -202,10 +202,20 @@ func thinkingOffUnlessAsked(body []byte) []byte {
 var anthropicModel = regexp.MustCompile(`(?i)(?:^|[/.:-])claude-`)
 
 // alwaysThinks is a vendor refusing to turn a model's thinking off: Z.ai's
-// GLM-5.3 answers 1210, "…always engages in thinking…", and DashScope's own
-// glm-5.3 answers "The value of the enable_thinking parameter is restricted
-// to True."
-var alwaysThinks = regexp.MustCompile(`(?i)always engages in thinking|thinking (?:can ?not|can't) be (?:disabled|turned off)|enable_thinking[^"]{0,60}restricted to true`)
+// GLM-5.3 answers 1210, "…always engages in thinking…", or, in Chinese
+// (ZCode's GLM-5.3-Flash), "该模型始终支持思考，不可关闭" (#699); DashScope's
+// own glm-5.3 answers "The value of the enable_thinking parameter is
+// restricted to True."
+var alwaysThinks = regexp.MustCompile(`(?i)always engages in thinking|thinking (?:can ?not|can't) be (?:disabled|turned off)|enable_thinking[^"]{0,60}restricted to true|始终(?:支持|开启|启用)?思考|思考[^"]{0,20}(?:不可|无法|不能)关闭`)
+
+// ThinksOnlyWhenAsked is a model that takes thinking turned off on the
+// Messages API whatever its levels: one of Anthropic's own (anthropicModel,
+// or a relay's opus-5.5). Another vendor's model there takes it off only
+// when none is among its levels; without it, it always thinks, and GLM-5.3
+// turns thinking disabled away (#699).
+func ThinksOnlyWhenAsked(model string) bool {
+	return anthropicModel.MatchString(model) || claudeVersion.MatchString(strings.ToLower(model))
+}
 
 // withoutThinkingOff is body with its thinking left to the model, when it
 // says thinking is off; false when it doesn't.
diff --git a/internal/gateway/gateway_test.go b/internal/gateway/gateway_test.go
index 99d18531..c6f240c5 100644
--- a/internal/gateway/gateway_test.go
+++ b/internal/gateway/gateway_test.go
@@ -589,6 +589,24 @@ func TestAnthropicPassthroughDashScopeThinkingRestricted(t *testing.T) {
 	}
 }
 
+// #699: ZCode's GLM-5.3-Flash refuses thinking turned off in Chinese
+// ("该模型始终支持思考，不可关闭"), which Pi's off sent and the agent got as
+// a 400: asked again with thinking left to the model, once
+func TestAnthropicPassthroughAlwaysThinksInChinese(t *testing.T) {
+	f := &fake{t: t, ctype: "application/json", reply: `{"id":"msg","type":"message","content":[]}`}
+	f.refuse = func(b []byte) (int, string) {
+		if bytes.Contains(b, []byte(`"disabled"`)) {
+			return 400, `{"type":"error","error":{"type":"invalid_request_error","message":"该模型始终支持思考，不可关闭"}}`
+		}
+		return 0, ""
+	}
+	setup(t, provider.Anthropic, f)
+	code, body := post(t, "/v1/messages", `{"model":"m1","max_tokens":5,"messages":[{"role":"user","content":"title?"}],"thinking":{"type":"disabled"}}`)
+	if code != 200 || f.calls != 2 || bytes.Contains(f.got, []byte("thinking")) || !bytes.Contains(f.got, []byte(`"title?"`)) {
+		t.Fatalf("%d %s after %d calls, last sent %s", code, body, f.calls, f.got)
+	}
+}
+
 func TestOpenRouterMandatoryReasoningRetries(t *testing.T) {
 	const refusal = `{"error":{"message":"OpenRouter: Reasoning is mandatory for this endpoint and cannot be disabled.","type":"invalid_request_error"},"type":"error"}`
 	for _, tc := range []struct {