diff --git a/config/config.example.json b/config/config.example.json
index adae6f05c..984eca6e2 100644
--- a/config/config.example.json
+++ b/config/config.example.json
@@ -6,7 +6,8 @@
       "model_name": "gpt4",
       "max_tokens": 8192,
       "temperature": 0.7,
-      "max_tool_iterations": 20
+      "max_tool_iterations": 20,
+      "thinking_level": ""
     }
   },
   "model_list": [
diff --git a/pkg/agent/loop.go b/pkg/agent/loop.go
index d111c10a4..f2d7afa28 100644
--- a/pkg/agent/loop.go
+++ b/pkg/agent/loop.go
@@ -779,7 +779,12 @@ func (al *AgentLoop) runLLMIteration(
 		// parseThinkingLevel guarantees ThinkingOff for empty/unknown values,
 		// so checking != ThinkingOff is sufficient.
 		if agent.ThinkingLevel != ThinkingOff {
-			llmOpts["thinking_level"] = string(agent.ThinkingLevel)
+			if tc, ok := agent.Provider.(providers.ThinkingCapable); ok && tc.SupportsThinking() {
+				llmOpts["thinking_level"] = string(agent.ThinkingLevel)
+			} else {
+				logger.WarnCF("agent", "thinking_level is set but current provider does not support it, ignoring",
+					map[string]any{"agent_id": agent.ID, "thinking_level": string(agent.ThinkingLevel)})
+			}
 		}
 
 		callLLM := func() (*providers.LLMResponse, error) {
diff --git a/pkg/providers/anthropic/provider.go b/pkg/providers/anthropic/provider.go
index b98cdc622..1b250b9b4 100644
--- a/pkg/providers/anthropic/provider.go
+++ b/pkg/providers/anthropic/provider.go
@@ -31,6 +31,9 @@ type Provider struct {
 	baseURL     string
 }
 
+// SupportsThinking implements providers.ThinkingCapable.
+func (p *Provider) SupportsThinking() bool { return true }
+
 func NewProvider(token string) *Provider {
 	return NewProviderWithBaseURL(token, "")
 }
@@ -202,6 +205,9 @@ func buildParams(
 func applyThinkingConfig(params *anthropic.MessageNewParams, level string) {
 	// Anthropic API rejects requests with temperature set alongside thinking.
 	// Reset to zero value (omitted from JSON serialization).
+	if params.Temperature.Valid() {
+		log.Printf("anthropic: temperature cleared because thinking is enabled (level=%s)", level)
+	}
 	params.Temperature = anthropic.MessageNewParams{}.Temperature
 
 	if level == "adaptive" {
@@ -220,22 +226,34 @@ func applyThinkingConfig(params *anthropic.MessageNewParams, level string) {
 
 	// budget_tokens must be < max_tokens; clamp to respect user's max_tokens setting.
 	if budget >= params.MaxTokens {
+		log.Printf("anthropic: budget_tokens (%d) clamped to %d (max_tokens-1)", budget, params.MaxTokens-1)
 		budget = params.MaxTokens - 1
+	} else if budget > params.MaxTokens*80/100 {
+		log.Printf("anthropic: thinking budget (%d) exceeds 80%% of max_tokens (%d), output may be truncated",
+			budget, params.MaxTokens)
 	}
 	params.Thinking = anthropic.ThinkingConfigParamOfEnabled(budget)
 }
 
-// levelToBudget maps a thinking level string to budget_tokens for legacy models.
+// levelToBudget maps a thinking level to budget_tokens.
+// Values are based on Anthropic's recommendations and community best practices:
+//
+//	low    =  4,096  — simple reasoning, quick debugging (Claude Code "think")
+//	medium = 16,384  — Anthropic recommended sweet spot for most tasks
+//	high   = 32,000  — complex architecture, deep analysis (diminishing returns above this)
+//	xhigh  = 64,000  — extreme reasoning, research problems, benchmarks
+//
+// Note: For Claude 4.6+, prefer adaptive thinking over manual budget_tokens.
 func levelToBudget(level string) int {
 	switch level {
 	case "low":
 		return 4096
 	case "medium":
-		return 10000
+		return 16384
 	case "high":
 		return 32000
 	case "xhigh":
-		return 128000
+		return 64000
 	default:
 		return 0
 	}
diff --git a/pkg/providers/anthropic/thinking_test.go b/pkg/providers/anthropic/thinking_test.go
index 6f0dac5d9..e69a3869e 100644
--- a/pkg/providers/anthropic/thinking_test.go
+++ b/pkg/providers/anthropic/thinking_test.go
@@ -1,6 +1,7 @@
 package anthropicprovider
 
 import (
+	"encoding/json"
 	"testing"
 
 	"github.com/anthropics/anthropic-sdk-go"
@@ -33,9 +34,9 @@ func TestApplyThinkingConfig_BudgetLevels(t *testing.T) {
 		wantBudget int64
 	}{
 		{"low", 4096},
-		{"medium", 10000},
+		{"medium", 16384},
 		{"high", 32000},
-		{"xhigh", 128000},
+		{"xhigh", 64000},
 	}
 
 	for _, tt := range tests {
@@ -100,9 +101,9 @@ func TestLevelToBudget(t *testing.T) {
 		want  int
 	}{
 		{"low", "low", 4096},
-		{"medium", "medium", 10000},
+		{"medium", "medium", 16384},
 		{"high", "high", 32000},
-		{"xhigh", "xhigh", 128000},
+		{"xhigh", "xhigh", 64000},
 		{"off", "off", 0},
 		{"empty", "", 0},
 	}
@@ -135,8 +136,59 @@ func TestBuildParams_ThinkingClearsTemperature(t *testing.T) {
 	if params.Thinking.OfEnabled == nil {
 		t.Fatal("expected enabled thinking")
 	}
-	if params.Thinking.OfEnabled.BudgetTokens != 10000 {
-		t.Errorf("budget_tokens = %d, want 10000", params.Thinking.OfEnabled.BudgetTokens)
+	if params.Thinking.OfEnabled.BudgetTokens != 16384 {
+		t.Errorf("budget_tokens = %d, want 16384", params.Thinking.OfEnabled.BudgetTokens)
+	}
+}
+
+// unmarshalBlocks constructs []ContentBlockUnion via JSON round-trip so that
+// the internal JSON.raw field is populated (required by AsText/AsThinking).
+func unmarshalBlocks(t *testing.T, jsonStr string) []anthropic.ContentBlockUnion {
+	t.Helper()
+	var blocks []anthropic.ContentBlockUnion
+	if err := json.Unmarshal([]byte(jsonStr), &blocks); err != nil {
+		t.Fatalf("unmarshalBlocks: %v", err)
+	}
+	return blocks
+}
+
+func TestParseResponse_ThinkingBlock(t *testing.T) {
+	resp := &anthropic.Message{
+		Content: unmarshalBlocks(t, `[
+			{"type":"thinking","thinking":"Let me reason step by step...","signature":"sig"},
+			{"type":"text","text":"The answer is 42."}
+		]`),
+		StopReason: anthropic.StopReasonEndTurn,
+	}
+
+	result := parseResponse(resp)
+
+	if result.Reasoning != "Let me reason step by step..." {
+		t.Errorf("Reasoning = %q, want thinking content", result.Reasoning)
+	}
+	if result.Content != "The answer is 42." {
+		t.Errorf("Content = %q, want text content", result.Content)
+	}
+	if result.FinishReason != "stop" {
+		t.Errorf("FinishReason = %q, want stop", result.FinishReason)
+	}
+}
+
+func TestParseResponse_NoThinkingBlock(t *testing.T) {
+	resp := &anthropic.Message{
+		Content: unmarshalBlocks(t, `[
+			{"type":"text","text":"Just a normal response."}
+		]`),
+		StopReason: anthropic.StopReasonEndTurn,
+	}
+
+	result := parseResponse(resp)
+
+	if result.Reasoning != "" {
+		t.Errorf("Reasoning = %q, want empty", result.Reasoning)
+	}
+	if result.Content != "Just a normal response." {
+		t.Errorf("Content = %q, want text content", result.Content)
 	}
 }
 
diff --git a/pkg/providers/types.go b/pkg/providers/types.go
index f0c168bc6..68bbd1e65 100644
--- a/pkg/providers/types.go
+++ b/pkg/providers/types.go
@@ -37,6 +37,13 @@ type StatefulProvider interface {
 	Close()
 }
 
+// ThinkingCapable is an optional interface for providers that support
+// extended thinking (e.g. Anthropic). Used by the agent loop to warn
+// when thinking_level is configured but the active provider cannot use it.
+type ThinkingCapable interface {
+	SupportsThinking() bool
+}
+
 // FailoverReason classifies why an LLM request failed for fallback decisions.
 type FailoverReason string