@@ -289,11 +289,12 @@ pub fn list_all_models(conn: &Connection) -> Result<Vec<serde_json::Value>> {
     let mut stmt = conn.prepare(
         "SELECT id, provider, model, currency, input_per_m, output_per_m, \
                 context_length, max_output, vision, cache_hit_input_per_m, \
-                peak_input_per_m, peak_output_per_m, peak_cache_hit_input_per_m, updated_at \
+                cache_write_input_per_m, peak_input_per_m, peak_output_per_m, \
+                peak_cache_hit_input_per_m, updated_at \
          FROM models ORDER BY provider, model",
     )?;
     let rows = stmt.query_map([], |r| {
-        let updated_at: String = r.get(13)?;
+        let updated_at: String = r.get(14)?;
         Ok(serde_json::json!({
             "id": r.get::<_, i64>(0)?,
             "provider": r.get::<_, String>(1)?,
@@ -305,9 +306,10 @@ pub fn list_all_models(conn: &Connection) -> Result<Vec<serde_json::Value>> {
             "max_output": r.get::<_, i64>(7)?,
             "vision": r.get::<_, i64>(8)?,
             "cache_hit_input_per_m": r.get::<_, f64>(9)?,
-            "peak_input_per_m": r.get::<_, f64>(10)?,
-            "peak_output_per_m": r.get::<_, f64>(11)?,
-            "peak_cache_hit_input_per_m": r.get::<_, f64>(12)?,
+            "cache_write_input_per_m": r.get::<_, f64>(10)?,
+            "peak_input_per_m": r.get::<_, f64>(11)?,
+            "peak_output_per_m": r.get::<_, f64>(12)?,
+            "peak_cache_hit_input_per_m": r.get::<_, f64>(13)?,
             "updated_at": utc_iso(&updated_at),
         }))
     })?;
 
@@ -2,8 +2,8 @@
 //!
 //! - GET    /api/admin/models：全部模型列表（含 id / 所有字段，供管理表格）
 //! - POST   /api/admin/models：新增 {provider, model, currency, input_per_m, output_per_m,
-//!   context_length?, max_output?, vision?, cache_hit_input_per_m?, peak_input_per_m?,
-//!   peak_output_per_m?, peak_cache_hit_input_per_m?}
+//!   context_length?, max_output?, vision?, cache_hit_input_per_m?, cache_write_input_per_m?,
+//!   peak_input_per_m?, peak_output_per_m?, peak_cache_hit_input_per_m?}
 //!   校验：provider/model 非空、价格 ≥ 0、currency ∈ {USD, CNY}、(provider, model) 唯一冲突 409
 //! - PATCH  /api/admin/models/:id：部分更新（任意字段）
 //! - DELETE /api/admin/models/:id：直接删除（只删目录行、**不改写 keys**；删除的后果由路由入口的
@@ -40,6 +40,9 @@ pub struct ModelCreate {
     pub vision: i64,
     #[serde(default)]
     pub cache_hit_input_per_m: f64,
+    /// 缓存写入（cache-creation）输入价（#433；缺省 0 = 缓存写入不收费）
+    #[serde(default)]
+    pub cache_write_input_per_m: f64,
     /// 高峰时段价（rant 2026-08-20T11:58:40；缺省 0 = 不启用高峰计费）
     #[serde(default)]
     pub peak_input_per_m: f64,
@@ -65,19 +68,21 @@ pub struct ModelPatch {
     pub max_output: Option<i64>,
     pub vision: Option<i64>,
     pub cache_hit_input_per_m: Option<f64>,
+    pub cache_write_input_per_m: Option<f64>,
     pub peak_input_per_m: Option<f64>,
     pub peak_output_per_m: Option<f64>,
     pub peak_cache_hit_input_per_m: Option<f64>,
 }
 
 /// 校验通用字段：价格 ≥ 0、vision ∈ {0,1}、currency 枚举
-#[allow(clippy::too_many_arguments)] // 价格字段平铺（6 个价格 + currency + vision），保持与请求结构体同序
+#[allow(clippy::too_many_arguments)] // 价格字段平铺（7 个价格 + currency + vision），保持与请求结构体同序
 fn validate_common(
     currency: &str,
     input_per_m: f64,
     output_per_m: f64,
     vision: i64,
     cache_hit_input_per_m: f64,
+    cache_write_input_per_m: f64,
     peak_input_per_m: f64,
     peak_output_per_m: f64,
     peak_cache_hit_input_per_m: f64,
@@ -91,6 +96,7 @@ fn validate_common(
     if input_per_m < 0.0
         || output_per_m < 0.0
         || cache_hit_input_per_m < 0.0
+        || cache_write_input_per_m < 0.0
         || peak_input_per_m < 0.0
         || peak_output_per_m < 0.0
         || peak_cache_hit_input_per_m < 0.0
@@ -122,18 +128,19 @@ fn row_to_json(r: &rusqlite::Row) -> rusqlite::Result<serde_json::Value> {
         "max_output": r.get::<_, i64>(7)?,
         "vision": r.get::<_, i64>(8)?,
         "cache_hit_input_per_m": r.get::<_, f64>(9)?,
-        "peak_input_per_m": r.get::<_, f64>(10)?,
-        "peak_output_per_m": r.get::<_, f64>(11)?,
-        "peak_cache_hit_input_per_m": r.get::<_, f64>(12)?,
-        "updated_at": crate::dao::utc_iso(&r.get::<_, String>(13)?),
-        "context_window": r.get::<_, i64>(14)?,
+        "cache_write_input_per_m": r.get::<_, f64>(10)?,
+        "peak_input_per_m": r.get::<_, f64>(11)?,
+        "peak_output_per_m": r.get::<_, f64>(12)?,
+        "peak_cache_hit_input_per_m": r.get::<_, f64>(13)?,
+        "updated_at": crate::dao::utc_iso(&r.get::<_, String>(14)?),
+        "context_window": r.get::<_, i64>(15)?,
     }))
 }
 
 const ROW_SELECT: &str = "SELECT id, provider, model, currency, input_per_m, output_per_m, \
                 context_length, max_output, vision, cache_hit_input_per_m, \
-                peak_input_per_m, peak_output_per_m, peak_cache_hit_input_per_m, \
-                updated_at, context_window \
+                cache_write_input_per_m, peak_input_per_m, peak_output_per_m, \
+                peak_cache_hit_input_per_m, updated_at, context_window \
          FROM models ";
 
 /// GET /api/admin/models：全部模型（管理表格数据源）
@@ -168,6 +175,7 @@ pub async fn create(
         req.output_per_m,
         req.vision,
         req.cache_hit_input_per_m,
+        req.cache_write_input_per_m,
         req.peak_input_per_m,
         req.peak_output_per_m,
         req.peak_cache_hit_input_per_m,
@@ -190,9 +198,9 @@ pub async fn create(
         .query_row(
             "INSERT INTO models (provider, model, currency, input_per_m, output_per_m, \
                     context_length, max_output, vision, cache_hit_input_per_m, \
-                    peak_input_per_m, peak_output_per_m, peak_cache_hit_input_per_m, \
-                    context_window, updated_at) \
-             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, ?10, ?11, ?12, ?13, datetime('now')) \
+                    cache_write_input_per_m, peak_input_per_m, peak_output_per_m, \
+                    peak_cache_hit_input_per_m, context_window, updated_at) \
+             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, ?10, ?11, ?12, ?13, ?14, datetime('now')) \
              RETURNING id",
             params![
                 provider,
@@ -204,6 +212,7 @@ pub async fn create(
                 req.max_output,
                 req.vision,
                 req.cache_hit_input_per_m,
+                req.cache_write_input_per_m,
                 req.peak_input_per_m,
                 req.peak_output_per_m,
                 req.peak_cache_hit_input_per_m,
@@ -242,6 +251,7 @@ pub async fn patch(
         ));
     }
     // 部分更新：先取当前值，合并后校验 + 写回（简单且避免逐列拼接 SQL）
+    #[allow(clippy::type_complexity)] // 价格平铺（13 元组），同 validate_common
     let cur: (
         String,
         String,
@@ -255,11 +265,13 @@ pub async fn patch(
         f64,
         f64,
         f64,
+        f64,
     ) = conn
         .query_row(
             "SELECT provider, model, currency, input_per_m, output_per_m, \
                     context_length, max_output, vision, cache_hit_input_per_m, \
-                    peak_input_per_m, peak_output_per_m, peak_cache_hit_input_per_m \
+                    cache_write_input_per_m, peak_input_per_m, peak_output_per_m, \
+                    peak_cache_hit_input_per_m \
              FROM models WHERE id = ?1",
             [id],
             |r| {
@@ -276,6 +288,7 @@ pub async fn patch(
                     r.get(9)?,
                     r.get(10)?,
                     r.get(11)?,
+                    r.get(12)?,
                 ))
             },
         )
@@ -289,9 +302,10 @@ pub async fn patch(
     let max_output = req.max_output.unwrap_or(cur.6);
     let vision = req.vision.unwrap_or(cur.7);
     let cache_hit_input_per_m = req.cache_hit_input_per_m.unwrap_or(cur.8);
-    let peak_input_per_m = req.peak_input_per_m.unwrap_or(cur.9);
-    let peak_output_per_m = req.peak_output_per_m.unwrap_or(cur.10);
-    let peak_cache_hit_input_per_m = req.peak_cache_hit_input_per_m.unwrap_or(cur.11);
+    let cache_write_input_per_m = req.cache_write_input_per_m.unwrap_or(cur.9);
+    let peak_input_per_m = req.peak_input_per_m.unwrap_or(cur.10);
+    let peak_output_per_m = req.peak_output_per_m.unwrap_or(cur.11);
+    let peak_cache_hit_input_per_m = req.peak_cache_hit_input_per_m.unwrap_or(cur.12);
     if provider.is_empty() || model.is_empty() {
         return Err((
             axum::http::StatusCode::BAD_REQUEST,
@@ -304,6 +318,7 @@ pub async fn patch(
         output_per_m,
         vision,
         cache_hit_input_per_m,
+        cache_write_input_per_m,
         peak_input_per_m,
         peak_output_per_m,
         peak_cache_hit_input_per_m,
@@ -328,10 +343,11 @@ pub async fn patch(
     conn.execute(
         "UPDATE models SET provider = ?1, model = ?2, currency = ?3, input_per_m = ?4, \
                 output_per_m = ?5, context_length = ?6, max_output = ?7, vision = ?8, \
-                cache_hit_input_per_m = ?9, peak_input_per_m = ?10, peak_output_per_m = ?11, \
-                peak_cache_hit_input_per_m = ?12, context_window = ?6, tiers_json = '[]', \
+                cache_hit_input_per_m = ?9, cache_write_input_per_m = ?10, \
+                peak_input_per_m = ?11, peak_output_per_m = ?12, \
+                peak_cache_hit_input_per_m = ?13, context_window = ?6, tiers_json = '[]', \
                 updated_at = datetime('now') \
-         WHERE id = ?13",
+         WHERE id = ?14",
         params![
             provider,
             model,
@@ -342,6 +358,7 @@ pub async fn patch(
             max_output,
             vision,
             cache_hit_input_per_m,
+            cache_write_input_per_m,
             peak_input_per_m,
             peak_output_per_m,
             peak_cache_hit_input_per_m,
 
@@ -1307,6 +1307,124 @@ mod tests {
         assert_eq!(s, StatusCode::BAD_REQUEST, "高峰缓存命中价负数仍应 400");
     }
 
+    /// #433 的 `cache_write_input_per_m` 必须与其余价格字段同权：管理接口能写、`GET` 能读回。
+    /// 七个价各取**互不相同**的值 —— 任何一处列序错位（SELECT 与 `r.get(i)` 不同步）都会让
+    /// 某个断言读到别人的价而立刻变红，这是本用例要挡的那类回归。
+    #[tokio::test]
+    async fn admin_models_cache_write_price_round_trip() {
+        let st = test_state("admcachewrite");
+        let admin_bearer = login_bearer(&st, "admin@aitokenpool.local", "admin1234").await;
+
+        // 带值创建 → POST 响应回读
+        let (s, body) = post(
+            st.clone(),
+            "/api/admin/models",
+            r#"{"provider":"cw","model":"cw-billed","currency":"USD","input_per_m":1.0,"output_per_m":2.0,"cache_hit_input_per_m":0.5,"cache_write_input_per_m":0.25,"peak_input_per_m":1.5,"peak_output_per_m":3.0,"peak_cache_hit_input_per_m":0.75}"#,
+            Some(&admin_bearer),
+        )
+        .await;
+        assert_eq!(s, StatusCode::OK, "带缓存写入价创建应 200: {body}");
+        let v: serde_json::Value = serde_json::from_str(&body).unwrap();
+        let id = v["id"].as_i64().unwrap();
+        assert_eq!(v["cache_write_input_per_m"], 0.25, "POST 回读: {body}");
+        assert_eq!(v["input_per_m"], 1.0, "兄弟价原位: {body}");
+        assert_eq!(v["output_per_m"], 2.0, "兄弟价原位: {body}");
+        assert_eq!(v["cache_hit_input_per_m"], 0.5, "兄弟价原位: {body}");
+        assert_eq!(v["peak_input_per_m"], 1.5, "兄弟价原位: {body}");
+        assert_eq!(v["peak_output_per_m"], 3.0, "兄弟价原位: {body}");
+        assert_eq!(v["peak_cache_hit_input_per_m"], 0.75, "兄弟价原位: {body}");
+
+        // → GET 列表回读（列表走 dao::list_all_models，与单行 SELECT 是两条必须同步的读取路径）
+        let (_, body) = get(st.clone(), "/api/admin/models", Some(&admin_bearer)).await;
+        let arr: serde_json::Value = serde_json::from_str(&body).unwrap();
+        {
+            let row = arr
+                .as_array()
+                .unwrap()
+                .iter()
+                .find(|m| m["id"] == id)
+                .expect("模型仍在");
+            assert_eq!(row["cache_write_input_per_m"], 0.25, "GET 回读: {body}");
+            assert_eq!(row["input_per_m"], 1.0, "GET 兄弟价原位: {body}");
+            assert_eq!(row["cache_hit_input_per_m"], 0.5, "GET 兄弟价原位: {body}");
+            assert_eq!(
+                row["peak_cache_hit_input_per_m"], 0.75,
+                "GET 兄弟价原位: {body}"
+            );
+        }
+
+        // 不带该字段创建 → 0（缺省＝不收费；刻意**不**回落 input_per_m，否则未选入的部署会被重算价）
+        let (s, body) = post(
+            st.clone(),
+            "/api/admin/models",
+            r#"{"provider":"cw","model":"cw-default","currency":"USD","input_per_m":8.0,"output_per_m":9.0}"#,
+            Some(&admin_bearer),
+        )
+        .await;
+        assert_eq!(s, StatusCode::OK, "缺该字段创建应 200: {body}");
+        let v: serde_json::Value = serde_json::from_str(&body).unwrap();
+        assert_eq!(v["cache_write_input_per_m"], 0.0, "缺省 0 而非 8.0: {body}");
+
+        // PATCH 只改该价 → 生效，且其余价一律不动
+        let (s, body) = patch(
+            st.clone(),
+            &format!("/api/admin/models/{id}"),
+            r#"{"cache_write_input_per_m":3.5}"#,
+            Some(&admin_bearer),
+        )
+        .await;
+        assert_eq!(s, StatusCode::OK, "PATCH 缓存写入价应 200: {body}");
+        let v: serde_json::Value = serde_json::from_str(&body).unwrap();
+        assert_eq!(v["cache_write_input_per_m"], 3.5, "PATCH 回读: {body}");
+        assert_eq!(v["input_per_m"], 1.0, "PATCH 不得动兄弟价: {body}");
+        assert_eq!(
+            v["cache_hit_input_per_m"], 0.5,
+            "PATCH 不得动命中价: {body}"
+        );
+        assert_eq!(
+            v["peak_cache_hit_input_per_m"], 0.75,
+            "PATCH 不得动高峰命中价: {body}"
+        );
+
+        // PATCH 负数 → 400，且库中该值不被写入
+        let (s, body) = patch(
+            st.clone(),
+            &format!("/api/admin/models/{id}"),
+            r#"{"cache_write_input_per_m":-2.0}"#,
+            Some(&admin_bearer),
+        )
+        .await;
+        assert_eq!(s, StatusCode::BAD_REQUEST, "负数缓存写入价应 400: {body}");
+        let (_, body) = get(st.clone(), "/api/admin/models", Some(&admin_bearer)).await;
+        let arr: serde_json::Value = serde_json::from_str(&body).unwrap();
+        let row = arr
+            .as_array()
+            .unwrap()
+            .iter()
+            .find(|m| m["id"] == id)
+            .expect("模型仍在");
+        assert_eq!(
+            row["cache_write_input_per_m"], 3.5,
+            "被拒的 PATCH 不得改库: {body}"
+        );
+
+        // 创建时负数同样被拒（同一校验函数，两条入口都要有牙）
+        let (s, body) = post(
+            st.clone(),
+            "/api/admin/models",
+            r#"{"provider":"cw","model":"cw-negative","currency":"USD","cache_write_input_per_m":-0.5}"#,
+            Some(&admin_bearer),
+        )
+        .await;
+        assert_eq!(
+            s,
+            StatusCode::BAD_REQUEST,
+            "创建负数缓存写入价应 400: {body}"
+        );
+        let (_, body) = get(st.clone(), "/api/admin/models", Some(&admin_bearer)).await;
+        assert!(!body.contains("cw-negative"), "被拒的创建不得落库: {body}");
+    }
+
     #[tokio::test]
     async fn market_models_include_new_fields() {
         // GET /api/models 响应补 context_length/max_output/vision/cache_hit_input_per_m