diff --git a/crates/goose-provider-types/src/formats/openai_responses.rs b/crates/goose-provider-types/src/formats/openai_responses.rs
index 71a06090a1..a172558a29 100644
--- a/crates/goose-provider-types/src/formats/openai_responses.rs
+++ b/crates/goose-provider-types/src/formats/openai_responses.rs
@@ -150,21 +150,22 @@ pub struct ResponseUsage {
 pub struct InputTokensDetails {
     #[serde(default)]
     pub cached_tokens: Option<i32>,
+    #[serde(default)]
+    pub cache_write_tokens: Option<i32>,
 }
 
 impl ResponseUsage {
     fn to_usage(&self) -> Usage {
-        // input_tokens already includes cached tokens
-        let cached_tokens = self
-            .input_tokens_details
-            .as_ref()
-            .and_then(|d| d.cached_tokens);
+        // input_tokens already includes both cache reads and cache writes
+        let details = self.input_tokens_details.as_ref();
+        let cached_tokens = details.and_then(|d| d.cached_tokens);
+        let cache_write_tokens = details.and_then(|d| d.cache_write_tokens);
         Usage::new(
             Some(self.input_tokens),
             Some(self.output_tokens),
             Some(self.total_tokens),
         )
-        .with_cache_tokens(cached_tokens, None)
+        .with_cache_tokens(cached_tokens, cache_write_tokens)
     }
 }
 
diff --git a/crates/goose-provider-types/tests/openai_responses_cache_usage.rs b/crates/goose-provider-types/tests/openai_responses_cache_usage.rs
new file mode 100644
index 0000000000..2c818cdf65
--- /dev/null
+++ b/crates/goose-provider-types/tests/openai_responses_cache_usage.rs
@@ -0,0 +1,76 @@
+use futures::StreamExt;
+use goose_provider_types::conversation::token_usage::Usage;
+use goose_provider_types::formats::openai_responses::{
+    get_responses_usage, responses_api_to_streaming_message, ResponsesApiResponse,
+};
+use serde_json::json;
+
+async fn final_stream_usage(lines: Vec<String>) -> Usage {
+    let messages =
+        responses_api_to_streaming_message(tokio_stream::iter(lines.into_iter().map(Ok)));
+    futures::pin_mut!(messages);
+
+    let mut usage = None;
+    while let Some(item) = messages.next().await {
+        let (_, maybe_usage) = item.expect("stream item should parse");
+        if maybe_usage.is_some() {
+            usage = maybe_usage;
+        }
+    }
+    usage.expect("stream should report usage").usage
+}
+
+#[tokio::test]
+async fn stream_completed_reports_cache_write_tokens() {
+    let usage = final_stream_usage(vec![
+        r#"data: {"type":"response.created","sequence_number":1,"response":{"id":"resp_1","object":"response","created_at":1737368310,"status":"in_progress","model":"gpt-6-luna","output":[]}}"#.to_string(),
+        r#"data: {"type":"response.completed","sequence_number":2,"response":{"id":"resp_1","object":"response","created_at":1737368310,"status":"completed","model":"gpt-6-luna","output":[],"usage":{"input_tokens":5457,"input_tokens_details":{"cache_write_tokens":5454,"cached_tokens":0},"output_tokens":5,"output_tokens_details":{"reasoning_tokens":0},"total_tokens":5462}}}"#.to_string(),
+        "data: [DONE]".to_string(),
+    ])
+    .await;
+
+    assert_eq!(usage.input_tokens, Some(5457));
+    assert_eq!(usage.total_tokens, Some(5462));
+    assert_eq!(usage.cache_read_input_tokens, Some(0));
+    assert_eq!(usage.cache_write_input_tokens, Some(5454));
+}
+
+#[tokio::test]
+async fn stream_incomplete_reports_cache_write_tokens() {
+    let usage = final_stream_usage(vec![
+        r#"data: {"type":"response.created","sequence_number":1,"response":{"id":"resp_1","object":"response","created_at":1737368310,"status":"in_progress","model":"gpt-6-luna","output":[]}}"#.to_string(),
+        r#"data: {"type":"response.incomplete","sequence_number":2,"response":{"id":"resp_1","object":"response","created_at":1737368310,"status":"incomplete","model":"gpt-6-luna","output":[],"incomplete_details":{"reason":"max_output_tokens"},"usage":{"input_tokens":10,"input_tokens_details":{"cache_write_tokens":6,"cached_tokens":3},"output_tokens":5,"total_tokens":15}}}"#.to_string(),
+        "data: [DONE]".to_string(),
+    ])
+    .await;
+
+    assert_eq!(usage.input_tokens, Some(10));
+    assert_eq!(usage.cache_read_input_tokens, Some(3));
+    assert_eq!(usage.cache_write_input_tokens, Some(6));
+}
+
+#[test]
+fn non_streaming_usage_reports_cache_read_and_write_tokens() {
+    let response: ResponsesApiResponse = serde_json::from_value(json!({
+        "id": "resp_1",
+        "object": "response",
+        "created_at": 1737368310,
+        "status": "completed",
+        "model": "gpt-6-luna",
+        "output": [],
+        "usage": {
+            "input_tokens": 2176,
+            "input_tokens_details": { "cache_write_tokens": 9, "cached_tokens": 2164 },
+            "output_tokens": 7,
+            "output_tokens_details": { "reasoning_tokens": 0 },
+            "total_tokens": 2183
+        }
+    }))
+    .unwrap();
+
+    let usage = get_responses_usage(&response);
+
+    assert_eq!(usage.input_tokens, Some(2176));
+    assert_eq!(usage.cache_read_input_tokens, Some(2164));
+    assert_eq!(usage.cache_write_input_tokens, Some(9));
+}