diff --git a/app/services/portfolio_analysis.py b/app/services/portfolio_analysis.py
index dfd3cb1..9de6383 100644
--- a/app/services/portfolio_analysis.py
+++ b/app/services/portfolio_analysis.py
@@ -344,7 +344,17 @@ async def analyse(
                     {"role": "system", "content": system},
                     {"role": "user", "content": user},
                 ],
-                max_tokens=2000,
+                # 4000 not 2000. Italian / Spanish / French / German
+                # output runs ~25-35% longer in tokens than English; on
+                # top of that DeepSeek-V4-flash bills its internal
+                # reasoning against the same budget. At 2000 we
+                # repeatedly hit finish_reason=length mid-sentence,
+                # which the reviewer agent then correctly flags as
+                # truncated and rejects — the user ends up looking at
+                # whatever stale row was last cached. 4000 leaves
+                # ample headroom; we only pay for tokens actually
+                # emitted, not the cap itself.
+                max_tokens=4000,
             )
             status = "ok"
             error_msg = None