diff --git a/internal/translator/kiro/claude/kiro_claude_request.go b/internal/translator/kiro/claude/kiro_claude_request.go
index 7012e644..0ad090ae 100644
--- a/internal/translator/kiro/claude/kiro_claude_request.go
+++ b/internal/translator/kiro/claude/kiro_claude_request.go
@@ -243,13 +243,11 @@ func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isA
 	// Process messages and build history
 	history, currentUserMsg, currentToolResults := processMessages(messages, modelID, origin)
 
-	// Build content with system prompt (only on first turn to avoid re-injection)
+	// Build content with system prompt.
+	// Keep thinking tags on subsequent turns so multi-turn Claude sessions
+	// continue to emit reasoning events.
 	if currentUserMsg != nil {
-		effectiveSystemPrompt := systemPrompt
-		if len(history) > 0 {
-			effectiveSystemPrompt = "" // Don't re-inject on subsequent turns
-		}
-		currentUserMsg.Content = buildFinalContent(currentUserMsg.Content, effectiveSystemPrompt, currentToolResults)
+		currentUserMsg.Content = buildFinalContent(currentUserMsg.Content, systemPrompt, currentToolResults)
 
 		// Deduplicate currentToolResults
 		currentToolResults = deduplicateToolResults(currentToolResults)
@@ -475,6 +473,15 @@ func IsThinkingEnabledWithHeaders(body []byte, headers http.Header) bool {
 		}
 	}
 
+	// Check model name directly for thinking hints.
+	// This enables thinking variants even when clients don't send explicit thinking fields.
+	model := strings.TrimSpace(gjson.GetBytes(body, "model").String())
+	modelLower := strings.ToLower(model)
+	if strings.Contains(modelLower, "thinking") || strings.Contains(modelLower, "-reason") {
+		log.Debugf("kiro: thinking mode enabled via model name hint: %s", model)
+		return true
+	}
+
 	log.Debugf("kiro: IsThinkingEnabled returning false (no thinking mode detected)")
 	return false
 }
diff --git a/internal/translator/kiro/openai/kiro_openai_request.go b/internal/translator/kiro/openai/kiro_openai_request.go
index 9515848f..474231b3 100644
--- a/internal/translator/kiro/openai/kiro_openai_request.go
+++ b/internal/translator/kiro/openai/kiro_openai_request.go
@@ -234,16 +234,16 @@ func BuildKiroPayloadFromOpenAI(openaiBody []byte, modelID, profileArn, origin s
 	// Kiro API supports official thinking/reasoning mode via <thinking_mode> tag.
 	// When set to "enabled", Kiro returns reasoning content as official reasoningContentEvent
 	// rather than inline <thinking> tags in assistantResponseEvent.
-	// We use a high max_thinking_length to allow extensive reasoning.
+	// Use a conservative thinking budget to reduce latency/cost spikes in long sessions.
 	if thinkingEnabled {
 		thinkingHint := `<thinking_mode>enabled</thinking_mode>
-<max_thinking_length>200000</max_thinking_length>`
+<max_thinking_length>16000</max_thinking_length>`
 		if systemPrompt != "" {
 			systemPrompt = thinkingHint + "\n\n" + systemPrompt
 		} else {
 			systemPrompt = thinkingHint
 		}
-		log.Debugf("kiro-openai: injected thinking prompt (official mode)")
+		log.Infof("kiro-openai: injected thinking prompt (official mode), has_tools: %v", len(kiroTools) > 0)
 	}
 
 	// Process messages and build history
@@ -831,7 +831,6 @@ func hasThinkingTagInBody(body []byte) bool {
 	return strings.Contains(bodyStr, "<thinking_mode>") || strings.Contains(bodyStr, "<max_thinking_length>")
 }
 
-
 // extractToolChoiceHint extracts tool_choice from OpenAI request and returns a system prompt hint.
 // OpenAI tool_choice values:
 // - "none": Don't use any tools