diff --git a/internal/modelgateway/README.en.md b/internal/modelgateway/README.en.md
index 5e7b49b3b..f83380f5f 100644
--- a/internal/modelgateway/README.en.md
+++ b/internal/modelgateway/README.en.md
@@ -25,8 +25,8 @@ SSE events flush continuously; error events inside HTTP200 are sanitized too. Re
 the upstream and a blocked downstream write. Each watcher has one request owner, cancellation and a join.
 
 The Anthropic surface targets Claude Code 2.1.303. POST messages accepts no query or exactly
-beta=true; the version is fixed to 2023-06-01 and beta tokens are limited to the nine in
-anthropic.go. Duplicate or unknown tokens fail. Other client headers are dropped; the CLI's
+beta=true; the version is fixed to 2023-06-01 and beta tokens are limited to the claudeBetaHeaders
+set in anthropic.go. Duplicate or unknown tokens fail. Other client headers are dropped; the CLI's
 browser-access marker is validated and discarded. GET models (no query or limit=1000) rechecks
 current authorization and returns only the frozen model without an upstream call, key decryption
 or inference-attempt record. Optional endpoints such as count_tokens remain explicitly unsupported.
diff --git a/internal/modelgateway/README.md b/internal/modelgateway/README.md
index f9bdd13ef..1e0394ce2 100644
--- a/internal/modelgateway/README.md
+++ b/internal/modelgateway/README.md
@@ -23,7 +23,7 @@
 每个授权观察任务归属于一个请求，取消后必须等待它退出。
 
 Anthropic 兼容面固定到 Claude Code 2.1.303：推理只接受 POST messages 的空 query 或
-精确 beta=true；协议版本固定 2023-06-01，beta token 仅允许 anthropic.go 列出的九项，
+精确 beta=true；协议版本固定 2023-06-01，beta token 仅允许 anthropic.go 中 claudeBetaHeaders 列出的值，
 重复或未知值拒绝。其他客户端头不转发；CLI 的 browser-access 标记只验证并丢弃。
 GET models（空 query 或 limit=1000）在重新核验当前授权后返回绑定的唯一模型，不请求上游、
 不解密 Key、不登记推理 attempt。count_tokens 等可选接口保持未实现并明确拒绝。
diff --git a/internal/modelgateway/anthropic_test.go b/internal/modelgateway/anthropic_test.go
index 9e2b883de..a0367fd5e 100644
--- a/internal/modelgateway/anthropic_test.go
+++ b/internal/modelgateway/anthropic_test.go
@@ -19,7 +19,8 @@ const pinnedClaudeBetas = "claude-code-20250219,interleaved-thinking-2025-05-14,
 // Each pinned CLI's exact first-request header for a Claude model is forwarded unchanged.
 func TestPinnedCLIClaudeModelBetas(t *testing.T) {
 	for name, betas := range map[string]string{
-		// Claude Code 2.1.303 through the pinned ACP adapter, with claude-haiku-5-5.
+		// Claude Code 2.1.303 run directly with the adapter's alias and model environment, with
+		// claude-haiku-5-5; the Agent SDK's own additions are covered separately below.
 		"claude": "interleaved-thinking-2025-05-14,thinking-token-count-2026-05-13,context-management-2025-06-27,prompt-caching-scope-2026-01-05,claude-code-20250219,per-turn-control-2026-07-01,mid-conversation-tool-changes-2026-07-01,effort-2025-11-24",
 		// OpenCode 1.19.3 through @ai-sdk/anthropic, with claude-haiku-5-5.
 		"opencode": "thinking-binding-controls-2026-08-01,structured-outputs-2025-11-13",
@@ -48,6 +49,30 @@ func TestPinnedCLIClaudeModelBetas(t *testing.T) {
 	}
 }
 
+// Agent SDK query creation enables these additional capabilities on the same pinned native CLI.
+func TestPinnedClaudeAgentSDKCapabilities(t *testing.T) {
+	betas := pinnedClaudeBetas + ",structured-outputs-2025-12-15,thinking-display-updates-2026-08-18"
+	var calls atomic.Int64
+	service, _ := anthropicFixture(t, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls.Add(1)
+		if r.Header.Get("Anthropic-Beta") != betas || r.URL.RawQuery != "beta=true" {
+			t.Error("SDK capabilities changed while forwarding")
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = io.WriteString(w, "{\"id\":\"m\",\"type\":\"message\",\"role\":\"assistant\",\"model\":\"fixture-model\",\"content\":[{\"type\":\"text\",\"text\":\"{}\"}],\"stop_reason\":\"end_turn\",\"usage\":{\"input_tokens\":3,\"output_tokens\":1}}")
+	}))
+	request := httptest.NewRequest(http.MethodPost, "/runtime/anthropic/v1/messages?beta=true", strings.NewReader("{\"model\":\"fixture-model\",\"messages\":[],\"max_tokens\":16}"))
+	request.Header.Set("Authorization", "Bearer "+strings.Repeat("r", 40))
+	request.Header.Set("Content-Type", "application/json")
+	request.Header.Set("Anthropic-Version", anthropicVersion)
+	request.Header.Set("Anthropic-Beta", betas)
+	response := httptest.NewRecorder()
+	service.ModelHandler(response, request)
+	if response.Code != 200 || calls.Load() != 1 {
+		t.Fatalf("pinned SDK request rejected: status=%d calls=%d", response.Code, calls.Load())
+	}
+}
+
 // anthropicFixture keeps the upstream key separate from the workload's short-lived token.
 func anthropicFixture(t *testing.T, handler http.Handler) (*Service, *fakePolicy) {
 	t.Helper()