diff --git a/crates/edit_prediction/src/edit_prediction.rs b/crates/edit_prediction/src/edit_prediction.rs
index ab88dbf2fa6a66..0f2a67c6a8a3aa 100644
--- a/crates/edit_prediction/src/edit_prediction.rs
+++ b/crates/edit_prediction/src/edit_prediction.rs
@@ -129,6 +129,7 @@ const GIT_CHANGED_FILE_SETS_COMMIT_LIMIT: usize = 100;
 const LAST_CHANGE_GROUPING_TIME: Duration = Duration::from_secs(1);
 const ZED_PREDICT_DATA_COLLECTION_CHOICE: &str = "zed_predict_data_collection_choice";
 const REJECT_REQUEST_DEBOUNCE: Duration = Duration::from_secs(15);
+const REQUEST_TIMEOUT_BACKOFF: Duration = Duration::from_secs(10);
 
 const EDIT_PREDICTION_SETTLED_TTL: Duration = Duration::from_secs(60 * 5);
 const EDIT_PREDICTION_SETTLED_QUIESCENCE: Duration = Duration::from_secs(10);
@@ -165,6 +166,7 @@ pub struct EditPredictionStore {
     update_required: bool,
     edit_prediction_model: EditPredictionModel,
     zeta2_raw_config: Option<Zeta2RawConfig>,
+    request_backoff_until: Option<Instant>,
     preferred_experiment: Option<String>,
     available_experiments: Vec<String>,
     pub mercury: Mercury,
@@ -925,6 +927,7 @@ impl EditPredictionStore {
             update_required: false,
             edit_prediction_model: EditPredictionModel::Zeta,
             zeta2_raw_config: Self::zeta2_raw_config_from_env(),
+            request_backoff_until: None,
             preferred_experiment: None,
             available_experiments: Vec::new(),
             mercury: Mercury::new(cx),
@@ -967,6 +970,27 @@ impl EditPredictionStore {
         self.zeta2_raw_config.as_ref()
     }
 
+    pub(crate) fn back_off_requests_after_timeout(&mut self, cx: &mut Context<Self>) {
+        self.request_backoff_until = Some(cx.background_executor().now() + REQUEST_TIMEOUT_BACKOFF);
+        log::info!(
+            "Backing off edit prediction requests for {:?} after Cloud timeout",
+            REQUEST_TIMEOUT_BACKOFF
+        );
+    }
+
+    fn request_backoff_active(&mut self, cx: &App) -> bool {
+        let Some(backoff_until) = self.request_backoff_until else {
+            return false;
+        };
+
+        if cx.background_executor().now() < backoff_until {
+            true
+        } else {
+            self.request_backoff_until = None;
+            false
+        }
+    }
+
     pub fn preferred_experiment(&self) -> Option<&str> {
         self.preferred_experiment.as_deref()
     }
@@ -2655,6 +2679,13 @@ impl EditPredictionStore {
             return Task::ready(Ok(None));
         }
 
+        if is_cloud_zeta && self.request_backoff_active(cx) {
+            log::debug!(
+                "Skipping Zeta edit prediction request while backing off after Cloud timeout"
+            );
+            return Task::ready(Ok(None));
+        }
+
         self.get_or_init_project(&project, cx);
         let project_state = self.projects.get(&project.entity_id()).unwrap();
         let stored_events = project_state.events(cx);
@@ -3029,6 +3060,9 @@ impl EditPredictionStore {
             let status = response.status();
             let mut body = String::new();
             response.body_mut().read_to_string(&mut body).await?;
+            if status == http_client::http::StatusCode::REQUEST_TIMEOUT {
+                return Err(anyhow::Error::new(CloudRequestTimeoutError));
+            }
             anyhow::bail!("Request failed with status: {status:?}\nBody: {body}");
         }
     }
@@ -3358,6 +3392,10 @@ pub struct ZedUpdateRequiredError {
     minimum_version: Version,
 }
 
+#[derive(Error, Debug)]
+#[error("Cloud request timed out")]
+pub(crate) struct CloudRequestTimeoutError;
+
 struct ZedPredictUpsell;
 
 fn is_upsell_dismissed(cx: &App) -> bool {
diff --git a/crates/edit_prediction/src/edit_prediction_tests.rs b/crates/edit_prediction/src/edit_prediction_tests.rs
index 52c6948cdc5a21..7d4a5f54c38df1 100644
--- a/crates/edit_prediction/src/edit_prediction_tests.rs
+++ b/crates/edit_prediction/src/edit_prediction_tests.rs
@@ -48,6 +48,7 @@ use zeta_prompt::ZetaPromptInput;
 use crate::{
     BufferEditPrediction, EDIT_PREDICTION_SETTLED_QUIESCENCE, EditPredictionId,
     EditPredictionJumpsFeatureFlag, EditPredictionStore, REJECT_REQUEST_DEBOUNCE,
+    REQUEST_TIMEOUT_BACKOFF,
 };
 
 #[gpui::test]
@@ -2066,6 +2067,63 @@ async fn test_jump_and_edit_throttles_are_independent(cx: &mut TestAppContext) {
     cx.run_until_parked();
 }
 
+#[gpui::test]
+async fn test_cloud_timeout_backs_off_zeta_requests(cx: &mut TestAppContext) {
+    let (ep_store, mut requests) = init_test_with_fake_client(cx);
+    let fs = FakeFs::new(cx.executor());
+    fs.insert_tree(
+        "/root",
+        json!({
+            "foo.md": "Hello!\nHow\nBye\n"
+        }),
+    )
+    .await;
+    let project = Project::test(fs, vec![path!("/root").as_ref()], cx).await;
+
+    let buffer = project
+        .update(cx, |project, cx| {
+            let path = project.find_project_path(path!("root/foo.md"), cx).unwrap();
+            project.open_buffer(path, cx)
+        })
+        .await
+        .unwrap();
+    let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot());
+    let position = snapshot.anchor_before(language::Point::new(1, 3));
+
+    ep_store.update(cx, |ep_store, cx| {
+        ep_store.register_project(&project, cx);
+        ep_store.register_buffer(&buffer, &project, cx);
+    });
+
+    ep_store.update(cx, |ep_store, cx| {
+        ep_store.refresh_prediction_from_buffer(project.clone(), buffer.clone(), position, cx);
+    });
+    let (_request, respond_tx) = requests.predict.next().await.unwrap();
+    respond_tx.send(request_timeout_response()).unwrap();
+    cx.run_until_parked();
+
+    ep_store.update(cx, |ep_store, cx| {
+        ep_store.refresh_prediction_from_buffer(project.clone(), buffer.clone(), position, cx);
+    });
+    cx.background_executor
+        .advance_clock(EditPredictionStore::THROTTLE_TIMEOUT);
+    cx.background_executor.run_until_parked();
+    cx.run_until_parked();
+    assert_no_predict_request_ready(&mut requests.predict);
+
+    cx.background_executor
+        .advance_clock(REQUEST_TIMEOUT_BACKOFF);
+    cx.background_executor.run_until_parked();
+    cx.run_until_parked();
+
+    ep_store.update(cx, |ep_store, cx| {
+        ep_store.refresh_prediction_from_buffer(project.clone(), buffer.clone(), position, cx);
+    });
+    let (_request, respond_tx) = requests.predict.next().await.unwrap();
+    respond_tx.send(empty_response()).unwrap();
+    cx.run_until_parked();
+}
+
 #[gpui::test]
 async fn test_same_frame_duplicate_requests_deduplicated(cx: &mut TestAppContext) {
     let (ep_store, mut requests) = init_test_with_fake_client(cx);
@@ -2546,6 +2604,18 @@ fn empty_response() -> PredictEditsV3Response {
     }
 }
 
+const REQUEST_TIMEOUT_RESPONSE_ID: &str = "__request_timeout__";
+
+fn request_timeout_response() -> PredictEditsV3Response {
+    PredictEditsV3Response {
+        request_id: REQUEST_TIMEOUT_RESPONSE_ID.to_string(),
+        editable_range: 0..0,
+        output: String::new(),
+        cursor_offset: None,
+        model_version: None,
+    }
+}
+
 fn prompt_from_request(request: &PredictEditsV3Request) -> String {
     zeta_prompt::format_zeta_prompt(&request.input, zeta_prompt::ZetaFormat::default())
         .expect("default zeta prompt formatting should succeed in edit prediction tests")
@@ -2659,9 +2729,20 @@ fn init_test_with_fake_client_and_legacy_data_collection(
                             let decompressed = zstd::decode_all(&buf[..]).unwrap();
                             let req = serde_json::from_slice(&decompressed).unwrap();
 
-                            let (res_tx, res_rx) = oneshot::channel();
+                            let (res_tx, res_rx) = oneshot::channel::<PredictEditsV3Response>();
                             predict_req_tx.unbounded_send((req, res_tx)).unwrap();
-                            serde_json::to_string(&res_rx.await?).unwrap()
+                            let response = res_rx.await?;
+                            if response.request_id == REQUEST_TIMEOUT_RESPONSE_ID {
+                                return Ok(Response::builder()
+                                    .status(http_client::http::StatusCode::REQUEST_TIMEOUT)
+                                    .body(
+                                        http_client::http::StatusCode::REQUEST_TIMEOUT
+                                            .as_str()
+                                            .into(),
+                                    )
+                                    .unwrap());
+                            }
+                            serde_json::to_string(&response).unwrap()
                         }
                         "/predict_edits/reject" => {
                             let mut buf = Vec::new();
diff --git a/crates/edit_prediction/src/zeta.rs b/crates/edit_prediction/src/zeta.rs
index ee2bcd62f04aa6..72463392dad977 100644
--- a/crates/edit_prediction/src/zeta.rs
+++ b/crates/edit_prediction/src/zeta.rs
@@ -1,7 +1,7 @@
 use crate::{
-    CurrentEditPrediction, DebugEvent, EditPredictionFinishedDebugEvent, EditPredictionId,
-    EditPredictionModelInput, EditPredictionStartedDebugEvent, EditPredictionStore,
-    ZedUpdateRequiredError, buffer_path_with_id_fallback,
+    CloudRequestTimeoutError, CurrentEditPrediction, DebugEvent, EditPredictionFinishedDebugEvent,
+    EditPredictionId, EditPredictionModelInput, EditPredictionStartedDebugEvent,
+    EditPredictionStore, ZedUpdateRequiredError, buffer_path_with_id_fallback,
     cursor_excerpt::{self, compute_cursor_excerpt, compute_syntax_ranges},
     data_collection::UncommittedDiffResult,
     prediction::EditPredictionResult,
@@ -491,6 +491,11 @@ fn handle_api_response<T>(
             Ok(data)
         }
         Err(err) => {
+            if err.is::<CloudRequestTimeoutError>() {
+                this.update(cx, |this, cx| this.back_off_requests_after_timeout(cx))
+                    .ok();
+            }
+
             if err.is::<ZedUpdateRequiredError>() {
                 cx.update(|cx| {
                     this.update(cx, |this, _cx| {