diff --git a/src/providers/openai.rs b/src/providers/openai.rs
index 8391a34..0f7c805 100644
--- a/src/providers/openai.rs
+++ b/src/providers/openai.rs
@@ -96,6 +96,7 @@ impl StreamingAccumulator {
     /// 跳过 total_tokens=0 的占位帧，避免覆盖真实值。
     fn set_usage(&mut self, usage: OpenAIUsage) {
         if usage.total_tokens > 0 {
+            usage.log_cache_diagnostics("stream_final_frame");
             self.usage = Some(usage);
         }
     }
@@ -724,6 +725,7 @@ impl OpenAIProvider {
                         })
                         .unwrap_or_default();
                     // 回退场景下也从非流式响应提取 usage
+                    openai_resp.usage.log_cache_diagnostics("non_streaming_fallback");
                     response.usage = Usage {
                         prompt_tokens: openai_resp.usage.prompt_tokens,
                         completion_tokens: openai_resp.usage.completion_tokens,
@@ -1106,6 +1108,19 @@ impl OpenAIUsage {
             })
             .unwrap_or(0)
     }
+
+    /// 诊断日志：记录 API 是否返回缓存字段及解析后的值（排查"缓存一直不命中"问题）。
+    /// field_present=false 说明 API/网关根本没返回缓存字段（中转服务剥离或模型不支持）。
+    fn log_cache_diagnostics(&self, source: &str) {
+        tracing::info!(
+            source = %source,
+            prompt_tokens = self.prompt_tokens,
+            cached_tokens = self.cached_tokens(),
+            deepseek_cache_field_present = self.prompt_cache_hit_tokens.is_some(),
+            openai_cache_details_present = self.prompt_tokens_details.is_some(),
+            "OpenAI usage cache diagnostics"
+        );
+    }
 }
 
 #[async_trait]
@@ -1263,11 +1278,14 @@ impl LLMProvider for OpenAIProvider {
             content,
             reasoning_content: openai_resp.choices[0].message.reasoning_content.clone(),
             tool_calls,
-            usage: Usage {
-                prompt_tokens: openai_resp.usage.prompt_tokens,
-                completion_tokens: openai_resp.usage.completion_tokens,
-                total_tokens: openai_resp.usage.total_tokens,
-                cached_tokens: openai_resp.usage.cached_tokens(),
+            usage: {
+                openai_resp.usage.log_cache_diagnostics("non_streaming");
+                Usage {
+                    prompt_tokens: openai_resp.usage.prompt_tokens,
+                    completion_tokens: openai_resp.usage.completion_tokens,
+                    total_tokens: openai_resp.usage.total_tokens,
+                    cached_tokens: openai_resp.usage.cached_tokens(),
+                }
             },
         })
     }
@@ -1571,6 +1589,45 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_usage_parses_deepseek_cache_fields() {
+        // DeepSeek 官方 API 的 usage 顶层缓存字段
+        let usage: OpenAIUsage = serde_json::from_value(json!({
+            "prompt_tokens": 345,
+            "completion_tokens": 564,
+            "total_tokens": 909,
+            "prompt_cache_hit_tokens": 128,
+            "prompt_cache_miss_tokens": 217
+        }))
+        .unwrap();
+
+        assert_eq!(usage.cached_tokens(), 128);
+        assert!(usage.prompt_cache_hit_tokens.is_some());
+    }
+
+    #[test]
+    fn test_usage_parses_openai_nested_cache_details() {
+        // OpenAI 兼容格式：嵌套在 prompt_tokens_details 中
+        let usage: OpenAIUsage = serde_json::from_value(json!({
+            "prompt_tokens": 1000,
+            "completion_tokens": 100,
+            "total_tokens": 1100,
+            "prompt_tokens_details": { "cached_tokens": 512 }
+        }))
+        .unwrap();
+
+        assert_eq!(usage.cached_tokens(), 512);
+
+        // 无缓存字段时回退为 0
+        let no_cache: OpenAIUsage = serde_json::from_value(json!({
+            "prompt_tokens": 10,
+            "completion_tokens": 5,
+            "total_tokens": 15
+        }))
+        .unwrap();
+        assert_eq!(no_cache.cached_tokens(), 0);
+    }
+
     #[test]
     fn test_openai_response_parses_json_tool_arguments() {
         let response: OpenAIResponse = serde_json::from_value(json!({