diff --git a/internal/runtime/executor/helps/usage_helpers.go b/internal/runtime/executor/helps/usage_helpers.go
index 6234aa1ed..c04c2c83c 100644
--- a/internal/runtime/executor/helps/usage_helpers.go
+++ b/internal/runtime/executor/helps/usage_helpers.go
@@ -460,7 +460,16 @@ func (b *StreamUsageBuffer) Observe(detail usage.Detail, ok bool) {
 	if b == nil || !ok {
 		return
 	}
-	b.detail = detail
+	responseServiceTier := strings.TrimSpace(detail.ResponseServiceTier)
+	if responseServiceTier == "" || hasNonZeroTokenUsage(detail) {
+		preservedTier := b.detail.ResponseServiceTier
+		b.detail = detail
+		if b.detail.ResponseServiceTier == "" {
+			b.detail.ResponseServiceTier = preservedTier
+		}
+	} else {
+		b.detail.ResponseServiceTier = responseServiceTier
+	}
 	b.ok = true
 }
 
@@ -482,12 +491,16 @@ func (b *StreamUsageBuffer) Detail() (usage.Detail, bool) {
 }
 
 func ParseCodexUsage(data []byte) (usage.Detail, bool) {
+	responseServiceTier := extractResponseServiceTier(data)
 	usageNode := gjson.ParseBytes(data).Get("response.usage")
 	if !hasOpenAIStyleUsageTokenFields(usageNode) {
-		return usage.Detail{}, false
+		if responseServiceTier == "" {
+			return usage.Detail{}, false
+		}
+		return usage.Detail{ResponseServiceTier: responseServiceTier}, true
 	}
 	detail := parseOpenAIStyleUsageNode(usageNode)
-	detail.ResponseServiceTier = extractResponseServiceTier(data)
+	detail.ResponseServiceTier = responseServiceTier
 	return detail, true
 }
 
@@ -500,12 +513,13 @@ func ParseCodexImageToolUsage(data []byte) (usage.Detail, bool) {
 }
 
 func ParseOpenAIUsage(data []byte) usage.Detail {
+	responseServiceTier := extractResponseServiceTier(data)
 	usageNode := gjson.ParseBytes(data).Get("usage")
 	if !hasOpenAIStyleUsageTokenFields(usageNode) {
-		return usage.Detail{}
+		return usage.Detail{ResponseServiceTier: responseServiceTier}
 	}
 	detail := parseOpenAIStyleUsageNode(usageNode)
-	detail.ResponseServiceTier = extractResponseServiceTier(data)
+	detail.ResponseServiceTier = responseServiceTier
 	return detail
 }
 
@@ -564,12 +578,16 @@ func ParseOpenAIStreamUsage(line []byte) (usage.Detail, bool) {
 	if len(payload) == 0 || !gjson.ValidBytes(payload) {
 		return usage.Detail{}, false
 	}
+	responseServiceTier := extractResponseServiceTier(payload)
 	usageNode := gjson.GetBytes(payload, "usage")
 	if !hasOpenAIStyleUsageTokenFields(usageNode) {
-		return usage.Detail{}, false
+		if responseServiceTier == "" {
+			return usage.Detail{}, false
+		}
+		return usage.Detail{ResponseServiceTier: responseServiceTier}, true
 	}
 	detail := parseOpenAIStyleUsageNode(usageNode)
-	detail.ResponseServiceTier = extractResponseServiceTier(payload)
+	detail.ResponseServiceTier = responseServiceTier
 	return detail, true
 }
 
diff --git a/internal/runtime/executor/helps/usage_helpers_test.go b/internal/runtime/executor/helps/usage_helpers_test.go
index b7107752f..0d1ff6713 100644
--- a/internal/runtime/executor/helps/usage_helpers_test.go
+++ b/internal/runtime/executor/helps/usage_helpers_test.go
@@ -88,6 +88,24 @@ func TestParseOpenAIUsageIgnoresNullUsage(t *testing.T) {
 	}
 }
 
+func TestParseOpenAIUsagePreservesResponseTierWithoutUsage(t *testing.T) {
+	t.Parallel()
+
+	detail := ParseOpenAIUsage([]byte(`{"service_tier":"default"}`))
+	if detail.ResponseServiceTier != "default" {
+		t.Fatalf("response service tier = %q, want default", detail.ResponseServiceTier)
+	}
+}
+
+func TestParseCodexUsagePreservesResponseTierWithoutUsage(t *testing.T) {
+	t.Parallel()
+
+	detail, ok := ParseCodexUsage([]byte(`{"response":{"service_tier":"default"}}`))
+	if !ok || detail.ResponseServiceTier != "default" {
+		t.Fatalf("ParseCodexUsage() = (%+v, %v), want response tier default", detail, ok)
+	}
+}
+
 func TestParseOpenAIStreamUsageIgnoresNullUsage(t *testing.T) {
 	line := []byte(`data: {"id":"chunk_1","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":null}],"usage":null}`)
 	if detail, ok := ParseOpenAIStreamUsage(line); ok {
@@ -145,6 +163,21 @@ func TestStreamUsageBufferKeepsLastUsage(t *testing.T) {
 	}
 }
 
+func TestStreamUsageBufferPreservesTierAcrossChunks(t *testing.T) {
+	t.Parallel()
+
+	var buffer StreamUsageBuffer
+	buffer.Observe(ParseOpenAIStreamUsage([]byte(`data: {"service_tier":"default"}`)))
+	buffer.Observe(ParseOpenAIStreamUsage([]byte(`data: {"usage":{"input_tokens":1,"output_tokens":1,"total_tokens":2}}`)))
+	detail, ok := buffer.Detail()
+	if !ok {
+		t.Fatal("Detail() ok = false, want true")
+	}
+	if detail.InputTokens != 1 || detail.OutputTokens != 1 || detail.ResponseServiceTier != "default" {
+		t.Fatalf("detail = %+v, want usage with response tier default", detail)
+	}
+}
+
 func TestStreamUsageBufferPreservesOnlyZeroUsage(t *testing.T) {
 	var buffer StreamUsageBuffer
 	buffer.Observe(usage.Detail{}, true)