diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 7f8832da005..76220ed3c67 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -326,6 +326,50 @@ def _enrich_http_exception_with_guardrail_context(
         detail.setdefault("guardrail_mode", event_hook)
 
 
+async def _patch_eval_spend_log(
+    prisma_client: Any,
+    request_id: str,
+    eval_info: Any,
+    _retried: bool = False,
+) -> None:
+    """Patch eval_information into an existing spend log row.
+
+    The DB-logger callback that writes spend logs runs concurrently with
+    async_post_call_success_hook. If the row isn't committed yet, we sleep
+    500 ms and retry once — covering the common timing gap without busy-looping.
+    """
+    import asyncio
+    import json as _json
+
+    try:
+        existing = await prisma_client.db.litellm_spendlogs.find_unique(
+            where={"request_id": request_id}
+        )
+        if existing is None:
+            if not _retried:
+                await asyncio.sleep(0.5)
+                await _patch_eval_spend_log(
+                    prisma_client, request_id, eval_info, _retried=True
+                )
+                return
+            verbose_proxy_logger.warning(
+                f"[Evals] Spend log row {request_id} not found after retry; eval_information not persisted"
+            )
+            return
+        meta = existing.metadata or {}
+        if isinstance(meta, str):
+            meta = _json.loads(meta)
+        meta["eval_information"] = eval_info
+        await prisma_client.db.litellm_spendlogs.update(
+            where={"request_id": request_id},
+            data={"metadata": _json.dumps(meta)},
+        )
+    except Exception as _e:
+        verbose_proxy_logger.warning(
+            f"[Evals] Failed to patch spend log {request_id}: {_e}"
+        )
+
+
 class ProxyLogging:
     """
     Logging/Custom Handlers for proxy.
@@ -2092,6 +2136,8 @@ class ProxyLogging:
 
             ############ Eval Gate [Beta] ##############################
             ############################################################
+            from litellm.proxy.proxy_server import prisma_client as _prisma_client
+
             try:
                 _agent_id = (data.get("metadata") or {}).get("agent_id") or data.get(
                     "agent_id"
@@ -2100,9 +2146,6 @@ class ProxyLogging:
                     from litellm.proxy.eval_management.eval_engine import (
                         run_evals_for_agent,
                     )
-                    from litellm.proxy.proxy_server import (
-                        prisma_client as _prisma_client,
-                    )
 
                     if _prisma_client is not None:
                         response = await run_evals_for_agent(
@@ -2112,10 +2155,37 @@ class ProxyLogging:
                             user_api_key_dict=user_api_key_dict,
                             prisma_client=_prisma_client,
                         )
+                        # Patch the spend log — eval runs after the spend logger fires,
+                        # so we update the row directly to make eval_information visible in Logs UI.
+                        # For A2A calls, the spend log request_id is the JSON-RPC id (data["id"]),
+                        # not the litellm_call_id UUID generated by common_request_processing.
+                        _eval_info = (data.get("metadata") or {}).get("eval_information")
+                        # For regular completions the spend log uses response.id (chatcmpl-XXX).
+                        # For A2A calls it uses the JSON-RPC id stored in data["id"].
+                        _request_id = (
+                            getattr(response, "id", None)
+                            or data.get("id")
+                            or data.get("litellm_call_id")
+                        )
+                        if _eval_info and _request_id:
+                            await _patch_eval_spend_log(
+                                _prisma_client, _request_id, _eval_info
+                            )
             except Exception as e:
                 from fastapi import HTTPException
 
                 if isinstance(e, HTTPException):
+                    # Even on block, patch the spend log so Logs UI shows the failed eval scores
+                    _eval_info = (data.get("metadata") or {}).get("eval_information")
+                    _request_id = (
+                        getattr(response, "id", None)
+                        or data.get("id")
+                        or data.get("litellm_call_id")
+                    )
+                    if _eval_info and _request_id and _prisma_client is not None:
+                        await _patch_eval_spend_log(
+                            _prisma_client, _request_id, _eval_info
+                        )
                     raise
                 verbose_proxy_logger.warning(
                     f"[Evals] Unexpected error in eval gate: {e}"