Open WebUI 0.12.5 Adds Conversation Execution Tracing, Fixes Duplicate Counting of Streaming Tokens

Open WebUI released version 0.12.5 on October 29, extending admin analytics to the execution process of individual conversations. The new Sessions view can filter by model, token range, sub-agents, and other criteria, and show saved replies along with the duration and results of tool calls. This lets operators investigate failure points beyond usage anomalies. Release notes

The update also fixes a statistics issue that affected cost analysis: some providers resend usage data with every streaming chunk. Under the previous behavior, those figures were repeatedly accumulated, inflating the token counts shown for replies and in admin analytics. The new version counts usage once per provider call; when a model makes multiple calls to use tools, it sums the usage from each call. This corrects how usage is recorded and does not imply that the model actually consumed fewer tokens. Release notes

The official documentation further explains what the analytics cover: Sessions include sub-agent, timer, and note conversations, and opening a session shows the execution trace for replies and tools. However, the spans in the list include idle time, so they should not be treated as inference latency. Detailed timings are stored only for new replies, meaning complete timing data cannot be reconstructed for older conversations. Analytics documentation

Access settings also affect deployment. Sessions requires ENABLE_ADMIN_CHAT_ACCESS to be enabled; related APIs return 403 when it is disabled. The documentation also explains that these admin controls determine what is visible through the product interface and API. Administrators can still access data through the database or server, so the setting should not be treated as a security boundary that isolates administrators. Configuration documentation

For engineering teams, this adds a new way to check agent workload and locate failures, but usage records from providers should still be cross-checked after deployment. Calls that do not report usage are not assigned values; when only some calls provide figures, usage is marked partial. Before comparing cost data from before and after the upgrade, confirm the model's Usage capability and the backend's reporting behavior to avoid mistaking a statistics correction for an efficiency improvement. Analytics documentation