fix: correct token usage accounting

Author: Algorithm5838

backend/open_webui/models/chat_messages.py

@@ -6,7 +6,7 @@
 from sqlalchemy import select, delete, func, cast, Integer, distinct
 from sqlalchemy.ext.asyncio import AsyncSession
 from open_webui.internal.db import Base, get_async_db_context
-from open_webui.utils.response import merge_usage, normalize_usage
+from open_webui.utils.response import normalize_usage
 from pydantic import BaseModel, ConfigDict
 from sqlalchemy import (
     JSON,
@@ -200,11 +200,14 @@ async def upsert_message(
                     existing.error = data.get('error')
                 if 'context_summary' in data or 'contextSummary' in data:
                     existing.context_summary = data.get('context_summary') or data.get('contextSummary')
-                # Extract and normalize usage
+                # Usage is monotonic per message; keep the largest snapshot so
+                # idempotent re-saves dedupe instead of double-counting.
                 usage = get_usage(data)
                 if usage:
-                    existing_usage = normalize_usage(existing.usage or {}) if existing.usage else {}
-                    existing.usage = existing_usage if usage == existing_usage else merge_usage(existing_usage, usage)
+                    new_total = normalize_usage(usage).get('total_tokens', 0)
+                    existing_total = normalize_usage(existing.usage or {}).get('total_tokens', 0)
+                    if existing.usage is None or new_total >= existing_total:
+                        existing.usage = usage
                 existing.updated_at = now
                 await db.commit()
                 await db.refresh(existing)

backend/open_webui/utils/middleware.py

@@ -4018,7 +4018,12 @@ def set_last_text(out, text):
                 else:
                     output = []
 
-            usage = None
+            # Continue re-sends prior context as prompt; input_tokens is cumulative across calls, not a double-count.
+            usage = (
+                normalize_usage(message['usage'])
+                if existing_output and message and message.get('usage')
+                else None
+            )
             prior_output = []
             last_response_id = None