diff --git a/changelog.d/fixes/6829-classify-404-model-not-found.md b/changelog.d/fixes/6829-classify-404-model-not-found.md
new file mode 100644
index 0000000000..77a741377c
--- /dev/null
+++ b/changelog.d/fixes/6829-classify-404-model-not-found.md
@@ -0,0 +1 @@
+- **fix(providers):** classify upstream `404` responses as `MODEL_NOT_FOUND` (model lockout) instead of a retryable provider error, stopping the retry storm when a single model is missing (#6829 — thanks @AndrianBalanescu).
diff --git a/open-sse/handlers/chatCore.ts b/open-sse/handlers/chatCore.ts
index 676292a4c5..8319e5ba80 100644
--- a/open-sse/handlers/chatCore.ts
+++ b/open-sse/handlers/chatCore.ts
@@ -3306,6 +3306,16 @@ export async function handleChatCore({
           console.warn(
             `[provider] Node ${errorConnectionId} project routing error (${statusCode}) — not banning`
           );
+        } else if (errorType === PROVIDER_ERROR_TYPES.MODEL_NOT_FOUND) {
+          // 404 — model/endpoint does not exist upstream. Lock the model so the
+          // retry/backoff loop stops hammering the dead endpoint (which would
+          // otherwise degenerate into a 429 rate-limit storm). Connection stays
+          // active since only the specific model is unavailable. (#6827)
+          const notFoundCooldownMs = COOLDOWN_MS.notFound;
+          lockModel(provider, errorConnectionId, currentModel, "model_not_found", notFoundCooldownMs);
+          console.warn(
+            `[provider] Node ${errorConnectionId} model not found (${statusCode}) for ${currentModel} - locking model for ${Math.ceil(notFoundCooldownMs / 1000)}s (connection stays active)`
+          );
         }
       } catch {
         // Best-effort state update; request flow should continue with fallback handling.
diff --git a/open-sse/services/errorClassifier.ts b/open-sse/services/errorClassifier.ts
index f68bebea3d..3d978fec4c 100644
--- a/open-sse/services/errorClassifier.ts
+++ b/open-sse/services/errorClassifier.ts
@@ -76,6 +76,7 @@ export const PROVIDER_ERROR_TYPES = {
   CONTEXT_OVERFLOW: "context_overflow",
   OAUTH_INVALID_TOKEN: "oauth_invalid_token",
   EMPTY_CONTENT: "empty_content",
+  MODEL_NOT_FOUND: "model_not_found",
 };
 
 export const CONTEXT_OVERFLOW_SIGNALS = [
@@ -144,6 +145,15 @@ export function classifyProviderError(
     return PROVIDER_ERROR_TYPES.RATE_LIMITED;
   }
 
+  // 404 — model or endpoint not found. Without classification the error
+  // falls through to `return null`, so no cooldown/lockout is applied and the
+  // retry/backoff loop keeps hammering the dead endpoint until the upstream
+  // rate-limits it (404 + 429 storm). Classify as MODEL_NOT_FOUND so the model
+  // gets locked via the cooldown layer and retries stop. (#6827)
+  if (statusCode === 404) {
+    return PROVIDER_ERROR_TYPES.MODEL_NOT_FOUND;
+  }
+
   if (statusCode === 401) {
     if (oauthInvalid) {
       return PROVIDER_ERROR_TYPES.OAUTH_INVALID_TOKEN;
diff --git a/tests/unit/error-classifier.test.ts b/tests/unit/error-classifier.test.ts
index c3a2ea1589..ef9b740955 100644
--- a/tests/unit/error-classifier.test.ts
+++ b/tests/unit/error-classifier.test.ts
@@ -120,3 +120,22 @@ test("classifyProviderError: OAuth provider 429 with daily quota signal => QUOTA
   );
   assert.equal(result, PROVIDER_ERROR_TYPES.QUOTA_EXHAUSTED);
 });
+
+// #6827 — 404 must be classified as MODEL_NOT_FOUND, not fall through to null.
+// Without this, no cooldown/lockout is applied and the retry loop keeps hitting
+// the dead endpoint until the upstream rate-limits it (404 + 429 storm).
+test("classifyProviderError: 404 => MODEL_NOT_FOUND", () => {
+  const result = classifyProviderError(404, {
+    error: { message: "model v0-1.5-md not found" },
+  });
+  assert.equal(result, PROVIDER_ERROR_TYPES.MODEL_NOT_FOUND);
+});
+
+test("classifyProviderError: 404 with provider => MODEL_NOT_FOUND", () => {
+  const result = classifyProviderError(
+    404,
+    { error: { message: "Not Found" } },
+    "v0-vercel"
+  );
+  assert.equal(result, PROVIDER_ERROR_TYPES.MODEL_NOT_FOUND);
+});