diff --git a/server-rs/crates/api-server/src/editor_screen_background_decision.rs b/server-rs/crates/api-server/src/editor_screen_background_decision.rs
index d2f59a7bd..1762e3e92 100644
--- a/server-rs/crates/api-server/src/editor_screen_background_decision.rs
+++ b/server-rs/crates/api-server/src/editor_screen_background_decision.rs
@@ -1,4 +1,4 @@
-use platform_llm::{LlmClient, LlmMessage, LlmTextRequest};
+use platform_llm::{LlmClient, LlmMessage, LlmResponseReasoningEffort, LlmTextRequest};
 use serde_json::json;
 use tracing::{info, warn};
 
@@ -151,11 +151,18 @@ pub(crate) async fn resolve_editor_screen_background_color(
             }
             None => LlmMessage::user(user_prompt.as_str()),
         };
+        // 预算要够推理模型（如 gpt-5-mini）先花几百 token 推理、再吐 JSON 答案；
+        // 实测 low 档推理约 320~384 token，取 768 留足余量。非推理模型遇 stop 提前结束，不会多花。
         let mut request = LlmTextRequest::new(vec![LlmMessage::system(system_prompt), user_message])
-            .with_max_tokens(96)
+            .with_max_tokens(768)
             .with_request_timeout_ms(EDITOR_SCREEN_BACKGROUND_DECISION_TIMEOUT_MS);
         if let Some(vision_model) = vision_model {
-            request = request.with_model(vision_model);
+            // 视觉模型 gpt-5-mini 是推理模型：走 Responses 协议并压到 low 推理档，
+            // 否则默认档会把预算全烧在推理上、返回空答案。
+            request = request
+                .with_model(vision_model)
+                .with_responses_api()
+                .with_response_reasoning_effort(LlmResponseReasoningEffort::Low);
         }
         match llm_client.request_text(request).await {
             Ok(response) => {
diff --git a/server-rs/crates/api-server/src/llm_model_routing.rs b/server-rs/crates/api-server/src/llm_model_routing.rs
index 16cd06539..49f8e6740 100644
--- a/server-rs/crates/api-server/src/llm_model_routing.rs
+++ b/server-rs/crates/api-server/src/llm_model_routing.rs
@@ -1,5 +1,6 @@
 pub(crate) const RPG_STORY_LLM_MODEL: &str = "doubao-seed-character-251128";
 pub(crate) const CREATION_TEMPLATE_LLM_MODEL: &str = "deepseek-v3-2-251201";
 pub(crate) const PUZZLE_LEVEL_NAME_VISION_LLM_MODEL: &str = "gpt-4o-mini";
-// 抠图背景色决策的视觉模型：只需看图选色，用低成本视觉模型即可。
-pub(crate) const EDITOR_SCREEN_BACKGROUND_VISION_LLM_MODEL: &str = "gpt-4o-mini";
+// 抠图背景色决策的视觉模型。gpt-5-mini 是推理模型，会先花若干 token 做推理，
+// 故决策请求的 max_tokens 需留足推理开销（见 editor_screen_background_decision）。
+pub(crate) const EDITOR_SCREEN_BACKGROUND_VISION_LLM_MODEL: &str = "gpt-5-mini";