diff --git a/providers/anthropic/anthropic.go b/providers/anthropic/anthropic.go
index 135ffe9c5..4b593408d 100644
--- a/providers/anthropic/anthropic.go
+++ b/providers/anthropic/anthropic.go
@@ -263,17 +263,19 @@ func (a languageModel) prepareParams(call fantasy.Call) (*anthropic.MessageNewPa
 		params.TopP = param.NewOpt(*call.TopP)
 	}
 
-	isThinking := false
-	var thinkingBudget int64
-	if providerOptions.Thinking != nil {
-		isThinking = true
-		thinkingBudget = providerOptions.Thinking.BudgetTokens
-	}
-	if isThinking {
-		if thinkingBudget == 0 {
+	switch {
+	case providerOptions.Effort != nil:
+		effort := *providerOptions.Effort
+		params.OutputConfig = anthropic.OutputConfigParam{
+			Effort: anthropic.OutputConfigEffort(effort),
+		}
+		adaptive := anthropic.NewThinkingConfigAdaptiveParam()
+		params.Thinking.OfAdaptive = &adaptive
+	case providerOptions.Thinking != nil:
+		if providerOptions.Thinking.BudgetTokens == 0 {
 			return nil, nil, &fantasy.Error{Title: "no budget", Message: "thinking requires budget"}
 		}
-		params.Thinking = anthropic.ThinkingConfigParamOfEnabled(thinkingBudget)
+		params.Thinking = anthropic.ThinkingConfigParamOfEnabled(providerOptions.Thinking.BudgetTokens)
 		if call.Temperature != nil {
 			params.Temperature = param.Opt[float64]{}
 			warnings = append(warnings, fantasy.CallWarning{
diff --git a/providers/anthropic/anthropic_test.go b/providers/anthropic/anthropic_test.go
index 4e808d405..897bc6e9f 100644
--- a/providers/anthropic/anthropic_test.go
+++ b/providers/anthropic/anthropic_test.go
@@ -1,8 +1,14 @@
 package anthropic
 
 import (
+	"context"
+	"encoding/json"
 	"errors"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
 	"testing"
+	"time"
 
 	"charm.land/fantasy"
 	"github.com/stretchr/testify/require"
@@ -401,3 +407,223 @@ func TestParseContextTooLargeError(t *testing.T) {
 		})
 	}
 }
+
+func TestParseOptions_Effort(t *testing.T) {
+	t.Parallel()
+
+	options, err := ParseOptions(map[string]any{
+		"send_reasoning":            true,
+		"thinking":                  map[string]any{"budget_tokens": int64(2048)},
+		"effort":                    "medium",
+		"disable_parallel_tool_use": true,
+	})
+	require.NoError(t, err)
+	require.NotNil(t, options.SendReasoning)
+	require.True(t, *options.SendReasoning)
+	require.NotNil(t, options.Thinking)
+	require.Equal(t, int64(2048), options.Thinking.BudgetTokens)
+	require.NotNil(t, options.Effort)
+	require.Equal(t, EffortMedium, *options.Effort)
+	require.NotNil(t, options.DisableParallelToolUse)
+	require.True(t, *options.DisableParallelToolUse)
+}
+
+func TestGenerate_SendsOutputConfigEffort(t *testing.T) {
+	t.Parallel()
+
+	server, calls := newAnthropicJSONServer(mockAnthropicGenerateResponse())
+	defer server.Close()
+
+	provider, err := New(
+		WithAPIKey("test-api-key"),
+		WithBaseURL(server.URL),
+	)
+	require.NoError(t, err)
+
+	model, err := provider.LanguageModel(context.Background(), "claude-sonnet-4-20250514")
+	require.NoError(t, err)
+
+	effort := EffortMedium
+	_, err = model.Generate(context.Background(), fantasy.Call{
+		Prompt: testPrompt(),
+		ProviderOptions: NewProviderOptions(&ProviderOptions{
+			Effort: &effort,
+		}),
+	})
+	require.NoError(t, err)
+
+	call := awaitAnthropicCall(t, calls)
+	require.Equal(t, "POST", call.method)
+	require.Equal(t, "/v1/messages", call.path)
+	requireAnthropicEffort(t, call.body, EffortMedium)
+}
+
+func TestStream_SendsOutputConfigEffort(t *testing.T) {
+	t.Parallel()
+
+	server, calls := newAnthropicStreamingServer([]string{
+		"event: message_start\n",
+		"data: {\"type\":\"message_start\",\"message\":{}}\n\n",
+		"event: message_stop\n",
+		"data: {\"type\":\"message_stop\"}\n\n",
+	})
+	defer server.Close()
+
+	provider, err := New(
+		WithAPIKey("test-api-key"),
+		WithBaseURL(server.URL),
+	)
+	require.NoError(t, err)
+
+	model, err := provider.LanguageModel(context.Background(), "claude-sonnet-4-20250514")
+	require.NoError(t, err)
+
+	effort := EffortHigh
+	stream, err := model.Stream(context.Background(), fantasy.Call{
+		Prompt: testPrompt(),
+		ProviderOptions: NewProviderOptions(&ProviderOptions{
+			Effort: &effort,
+		}),
+	})
+	require.NoError(t, err)
+
+	stream(func(fantasy.StreamPart) bool { return true })
+
+	call := awaitAnthropicCall(t, calls)
+	require.Equal(t, "POST", call.method)
+	require.Equal(t, "/v1/messages", call.path)
+	requireAnthropicEffort(t, call.body, EffortHigh)
+}
+
+type anthropicCall struct {
+	method string
+	path   string
+	body   map[string]any
+}
+
+func newAnthropicJSONServer(response map[string]any) (*httptest.Server, <-chan anthropicCall) {
+	calls := make(chan anthropicCall, 4)
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var body map[string]any
+		if r.Body != nil {
+			_ = json.NewDecoder(r.Body).Decode(&body)
+		}
+
+		calls <- anthropicCall{
+			method: r.Method,
+			path:   r.URL.Path,
+			body:   body,
+		}
+
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(response)
+	}))
+
+	return server, calls
+}
+
+func newAnthropicStreamingServer(chunks []string) (*httptest.Server, <-chan anthropicCall) {
+	calls := make(chan anthropicCall, 4)
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var body map[string]any
+		if r.Body != nil {
+			_ = json.NewDecoder(r.Body).Decode(&body)
+		}
+
+		calls <- anthropicCall{
+			method: r.Method,
+			path:   r.URL.Path,
+			body:   body,
+		}
+
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.Header().Set("Cache-Control", "no-cache")
+		w.Header().Set("Connection", "keep-alive")
+		w.WriteHeader(http.StatusOK)
+
+		for _, chunk := range chunks {
+			_, _ = fmt.Fprint(w, chunk)
+			if flusher, ok := w.(http.Flusher); ok {
+				flusher.Flush()
+			}
+		}
+	}))
+
+	return server, calls
+}
+
+func awaitAnthropicCall(t *testing.T, calls <-chan anthropicCall) anthropicCall {
+	t.Helper()
+
+	select {
+	case call := <-calls:
+		return call
+	case <-time.After(2 * time.Second):
+		t.Fatal("timed out waiting for Anthropic request")
+		return anthropicCall{}
+	}
+}
+
+func assertNoAnthropicCall(t *testing.T, calls <-chan anthropicCall) {
+	t.Helper()
+
+	select {
+	case call := <-calls:
+		t.Fatalf("expected no Anthropic API call, but got %s %s", call.method, call.path)
+	case <-time.After(200 * time.Millisecond):
+	}
+}
+
+func requireAnthropicEffort(t *testing.T, body map[string]any, expected Effort) {
+	t.Helper()
+
+	outputConfig, ok := body["output_config"].(map[string]any)
+	thinking, ok := body["thinking"].(map[string]any)
+	require.True(t, ok)
+	require.Equal(t, string(expected), outputConfig["effort"])
+	require.Equal(t, "adaptive", thinking["type"])
+}
+
+func testPrompt() fantasy.Prompt {
+	return fantasy.Prompt{
+		{
+			Role: fantasy.MessageRoleUser,
+			Content: []fantasy.MessagePart{
+				fantasy.TextPart{Text: "Hello"},
+			},
+		},
+	}
+}
+
+func mockAnthropicGenerateResponse() map[string]any {
+	return map[string]any{
+		"id":    "msg_01Test",
+		"type":  "message",
+		"role":  "assistant",
+		"model": "claude-sonnet-4-20250514",
+		"content": []any{
+			map[string]any{
+				"type": "text",
+				"text": "Hi there",
+			},
+		},
+		"stop_reason":   "end_turn",
+		"stop_sequence": "",
+		"usage": map[string]any{
+			"cache_creation": map[string]any{
+				"ephemeral_1h_input_tokens": 0,
+				"ephemeral_5m_input_tokens": 0,
+			},
+			"cache_creation_input_tokens": 0,
+			"cache_read_input_tokens":     0,
+			"input_tokens":                5,
+			"output_tokens":               2,
+			"server_tool_use": map[string]any{
+				"web_search_requests": 0,
+			},
+			"service_tier": "standard",
+		},
+	}
+}
diff --git a/providers/anthropic/provider_options.go b/providers/anthropic/provider_options.go
index 7c426f59a..ad6f6107b 100644
--- a/providers/anthropic/provider_options.go
+++ b/providers/anthropic/provider_options.go
@@ -7,6 +7,22 @@ import (
 	"charm.land/fantasy"
 )
 
+// Effort represents the output effort level for Anthropic models.
+//
+// This maps to Messages API `output_config.effort`.
+type Effort string
+
+const (
+	// EffortLow represents low output effort.
+	EffortLow Effort = "low"
+	// EffortMedium represents medium output effort.
+	EffortMedium Effort = "medium"
+	// EffortHigh represents high output effort.
+	EffortHigh Effort = "high"
+	// EffortMax represents maximum output effort.
+	EffortMax Effort = "max"
+)
+
 // Global type identifiers for Anthropic-specific provider data.
 const (
 	TypeProviderOptions         = Name + ".options"
@@ -43,6 +59,7 @@ func init() {
 type ProviderOptions struct {
 	SendReasoning          *bool                   `json:"send_reasoning"`
 	Thinking               *ThinkingProviderOption `json:"thinking"`
+	Effort                 *Effort                 `json:"effort"`
 	DisableParallelToolUse *bool                   `json:"disable_parallel_tool_use"`
 }