diff --git a/openless-all/app/src-tauri/src/polish.rs b/openless-all/app/src-tauri/src/polish.rs
index 4374d17e..1b152f90 100644
--- a/openless-all/app/src-tauri/src/polish.rs
+++ b/openless-all/app/src-tauri/src/polish.rs
@@ -35,7 +35,7 @@ pub struct OpenAICompatibleConfig {
     pub request_timeout_secs: u64,
     /// true = 让支持的 OpenAI-compatible provider 启用推理 / 思考；
     /// false = 按渠道级官方参数关闭或压低思考。不做模型白名单判断，
-    /// 具体模型兼容性交给 provider 处理。
+    /// 但 OpenAI 官方渠道会跳过已知不支持 reasoning_effort 的普通 chat 模型。
     pub thinking_enabled: bool,
 }
 
@@ -1509,13 +1509,20 @@ fn unix_now_secs() -> u64 {
 fn apply_openai_compatible_thinking_control(body: &mut Value, config: &OpenAICompatibleConfig) {
     match openai_compatible_thinking_control(&config.provider_id) {
         Some(ThinkingControl::ReasoningEffort) => {
-            // OpenAI Chat Completions 的 reasoning_effort 是渠道级请求字段。
-            // 关闭时统一压到 low，避免引入模型白名单；不支持该字段的模型由 provider 自行处理。
-            body["reasoning_effort"] = json!(if config.thinking_enabled {
-                "medium"
+            // OpenAI 官方 Chat Completions 只在推理模型族接受 reasoning_effort；
+            // 普通 chat 模型会直接 400。其它兼容渠道按渠道声明继续下发。
+            let effort = if config.provider_id.trim() == "openai" {
+                openai_chat_reasoning_effort(&config.model, config.thinking_enabled)
             } else {
-                "low"
-            });
+                Some(if config.thinking_enabled {
+                    "medium"
+                } else {
+                    "low"
+                })
+            };
+            if let Some(effort) = effort {
+                body["reasoning_effort"] = json!(effort);
+            }
         }
         Some(ThinkingControl::EnableThinking) => {
             body["enable_thinking"] = json!(config.thinking_enabled);
@@ -1554,6 +1561,28 @@ fn openai_compatible_thinking_control(provider_id: &str) -> Option<ThinkingContr
     }
 }
 
+fn openai_chat_reasoning_effort(model: &str, thinking_enabled: bool) -> Option<&'static str> {
+    let normalized = model
+        .trim()
+        .strip_prefix("openai/")
+        .unwrap_or_else(|| model.trim())
+        .to_ascii_lowercase();
+
+    if normalized.starts_with("gpt-5-pro") {
+        return Some("high");
+    }
+
+    if normalized.starts_with("o1")
+        || normalized.starts_with("o3")
+        || normalized.starts_with("o4")
+        || normalized.starts_with("gpt-5")
+    {
+        Some(if thinking_enabled { "medium" } else { "low" })
+    } else {
+        None
+    }
+}
+
 /// 把 working_languages + front_app 拼成 system prompt 头部前提：
 ///     # 上下文
 ///     用户的工作语言：…
@@ -2656,14 +2685,14 @@ mod tests {
     }
 
     #[test]
-    fn openai_chat_body_adds_reasoning_effort_for_openai_channel() {
+    fn openai_chat_body_adds_reasoning_effort_for_openai_reasoning_model() {
         let provider = OpenAICompatibleLLMProvider::new(
             OpenAICompatibleConfig::new(
                 "openai",
                 "OpenAI",
                 "https://api.openai.com/v1",
                 "k",
-                "any-model",
+                "gpt-5-mini",
             )
             .with_thinking_enabled(true),
         );
@@ -2673,6 +2702,49 @@ mod tests {
         assert_eq!(body["reasoning_effort"], "medium");
     }
 
+    #[test]
+    fn openai_chat_body_omits_reasoning_effort_for_non_reasoning_chat_models() {
+        for model in ["gpt-4o-mini", "gpt-4o", "gpt-4.1-nano"] {
+            let provider = OpenAICompatibleLLMProvider::new(
+                OpenAICompatibleConfig::new(
+                    "openai",
+                    "OpenAI",
+                    "https://api.openai.com/v1",
+                    "k",
+                    model,
+                )
+                .with_thinking_enabled(true),
+            );
+
+            let body = provider.chat_body(false, vec![json!({ "role": "user", "content": "hi" })]);
+
+            assert!(
+                body.get("reasoning_effort").is_none(),
+                "{model} must not receive reasoning_effort"
+            );
+        }
+    }
+
+    #[test]
+    fn openai_chat_body_uses_high_reasoning_effort_for_gpt_5_pro() {
+        for thinking_enabled in [false, true] {
+            let provider = OpenAICompatibleLLMProvider::new(
+                OpenAICompatibleConfig::new(
+                    "openai",
+                    "OpenAI",
+                    "https://api.openai.com/v1",
+                    "k",
+                    "gpt-5-pro",
+                )
+                .with_thinking_enabled(thinking_enabled),
+            );
+
+            let body = provider.chat_body(false, vec![json!({ "role": "user", "content": "hi" })]);
+
+            assert_eq!(body["reasoning_effort"], "high");
+        }
+    }
+
     #[test]
     fn openai_chat_body_lowers_reasoning_when_disabled_for_channel() {
         let provider = OpenAICompatibleLLMProvider::new(OpenAICompatibleConfig::new(
diff --git a/openless-all/app/src-tauri/src/types.rs b/openless-all/app/src-tauri/src/types.rs
index 1d9d925f..93b0ae08 100644
--- a/openless-all/app/src-tauri/src/types.rs
+++ b/openless-all/app/src-tauri/src/types.rs
@@ -532,7 +532,7 @@ pub struct UserPreferences {
     pub active_llm_provider: String, // "ark" | "openai" | ...
     /// LLM 思考模式开关。默认 false 以保持既有「尽量关闭思考」行为；
     /// Gemini 走原生 thinkingConfig，OpenAI-compatible 路径仅按 provider/channel
-    /// 下发官方渠道级字段，不用 prompt 注入，也不做模型白名单适配。详见 issue #402。
+    /// 下发官方渠道级字段；OpenAI 官方渠道会跳过普通 chat 模型不支持的字段。详见 issue #402。
     #[serde(default)]
     pub llm_thinking_enabled: bool,
     /// Windows/Linux 粘贴成功后是否恢复用户原剪贴板。默认 true 跟历史行为一致；
diff --git a/openless-all/app/src/lib/types.ts b/openless-all/app/src/lib/types.ts
index 245e7895..be033bf7 100644
--- a/openless-all/app/src/lib/types.ts
+++ b/openless-all/app/src/lib/types.ts
@@ -222,7 +222,7 @@ export interface UserPreferences {
   microphoneDeviceName: string;
   activeAsrProvider: string;
   activeLlmProvider: string;
-  /** LLM 思考模式开关。默认关闭，保持既有尽量关闭思考的行为。详见 issue #402。 */
+  /** LLM 思考模式开关。默认关闭；OpenAI 普通 chat 模型会跳过不支持的字段。详见 issue #402。 */
   llmThinkingEnabled: boolean;
   /** 仅 Windows/Linux：粘贴成功后是否恢复用户原剪贴板。默认 true。详见 issue #111。 */
   restoreClipboardAfterPaste: boolean;