diff --git a/Changelog.md b/Changelog.md
index ec18054..600f7f7 100644
--- a/Changelog.md
+++ b/Changelog.md
@@ -1,5 +1,9 @@
 # Changelog
 
+## Unreleased
+
+* Add `Error::Unavailable` to decouple service unavailability from 'queue full' 503 responses.
+
 ## 0.7.1
 
 * Add `Client::tokenize` and `Client::detokenize`. Thanks to @andreaskoepf
diff --git a/src/http.rs b/src/http.rs
index 1f91928..adefc02 100644
--- a/src/http.rs
+++ b/src/http.rs
@@ -160,9 +160,19 @@ async fn translate_http_error(response: reqwest::Response) -> Result<reqwest::Re
         // the API, but an intermediate Proxy like NGinx, so we can still forward the error
         // message.
         let body = response.text().await?;
+        // If the response is an error emitted by the API, this deserialization should succeed.
+        let api_error: Result<ApiError, _> = serde_json::from_str(&body);
         let translated_error = match status {
             StatusCode::TOO_MANY_REQUESTS => Error::TooManyRequests,
-            StatusCode::SERVICE_UNAVAILABLE => Error::Busy,
+            StatusCode::SERVICE_UNAVAILABLE => {
+                // Presence of `api_error` implies the error originated from the API itself (rather
+                // than the intermediate proxy) and so we can decode it as such.
+                if api_error.is_ok_and(|error| error.code == "QUEUE_FULL") {
+                    Error::Busy
+                } else {
+                    Error::Unavailable
+                }
+            }
             _ => Error::Http {
                 status: status.as_u16(),
                 body,
@@ -175,14 +185,14 @@ async fn translate_http_error(response: reqwest::Response) -> Result<reqwest::Re
 }
 
 /// We are only interested in the status codes of the API.
-#[derive(Deserialize)]
+#[derive(Deserialize, Debug)]
 struct ApiError<'a> {
     /// Unique string in capital letters emitted by the API to signal different kinds of errors in a
     /// finer granularity then the HTTP status codes alone would allow for.
     ///
     /// E.g. Differentiating between request rate limiting and parallel tasks limiting which both
     /// are 429 (the former is emitted by NGinx though).
-    _code: Cow<'a, str>,
+    code: Cow<'a, str>,
 }
 
 /// Errors returned by the Aleph Alpha Client
@@ -202,6 +212,12 @@ pub enum Error {
         welcome to retry your request any time."
     )]
     Busy,
+    /// The API itself is unavailable, most likely due to restart.
+    #[error(
+        "The service is currently unavailable. This is likely due to restart. Please try again \
+        later."
+    )]
+    Unavailable,
     #[error("No response received within given timeout: {0:?}")]
     ClientTimeout(Duration),
     /// An error on the Http Protocol level.
diff --git a/tests/unit.rs b/tests/unit.rs
index 28a9725..e6b2726 100644
--- a/tests/unit.rs
+++ b/tests/unit.rs
@@ -93,12 +93,12 @@ async fn detect_queue_full() {
     // Start a background HTTP server on a random local part
     let mock_server = MockServer::start().await;
 
-    let answer = r#"{
-        "error":"Sorry we had to reject your request because we could not guarantee to finish it in
-            a reasonable timeframe. This specific model is very busy at this moment. Try again later
-            or use another model.",
-        "code":"QUEUE_FULL"
-    }"#;
+    let answer = "{
+        \"error\":\"Sorry we had to reject your request because we could not guarantee to finish \
+            it in a reasonable timeframe. This specific model is very busy at this moment. Try \
+            again later or use another model.\",
+        \"code\":\"QUEUE_FULL\"
+    }";
     let body = r#"{
         "model": "luminous-base",
         "prompt": [{"type": "text", "data": "Hello,"}],
@@ -124,9 +124,47 @@ async fn detect_queue_full() {
         .await
         .unwrap_err();
 
+    // Then
     assert!(matches!(error, Error::Busy));
 }
 
+/// If the API is down, we want to detect this scenario and inform the user.
+#[tokio::test]
+async fn detect_service_unavailable() {
+    // Given
+
+    // Start a background HTTP server on a random local part
+    let mock_server = MockServer::start().await;
+
+    let answer = "No server is available to handle this request.";
+    let body = r#"{
+        "model": "luminous-base",
+        "prompt": [{"type": "text", "data": "Hello,"}],
+        "maximum_tokens": 1
+    }"#;
+
+    Mock::given(method("POST"))
+        .and(path("/complete"))
+        .and(header("Authorization", "Bearer dummy-token"))
+        .and(header("Content-Type", "application/json"))
+        .and(body_json_string(body))
+        .respond_with(ResponseTemplate::new(503).set_body_string(answer))
+        .mount(&mock_server)
+        .await;
+
+    // When
+    let task = TaskCompletion::from_text("Hello,", 1);
+    let model = "luminous-base";
+    let client = Client::with_base_url(mock_server.uri(), "dummy-token").unwrap();
+    let error = client
+        .output_of(&task.with_model(model), &How::default())
+        .await
+        .unwrap_err();
+
+    // Then
+    assert!(matches!(error, Error::Unavailable));
+}
+
 /// Should set `nice=true` in query URL in order to tell the server we do not need our result right
 /// now in a high stress situation.
 #[tokio::test]