Investigate a slow run

Updated 4 hours ago • October 18, 2026

Open a run with an unexpectedly long duration. Compare its overview with the trace before changing model or application settings.

Separate work from waiting

Look for an interrupt or a long quiet period. A person taking time to answer can make wall time much longer than the time spent doing work. Multiple model/tool calls can also overlap.

Open Trace and inspect the operations that span the slow part of the run. Their timestamps show whether calls happened in sequence or at the same time.

Check model timing

Time to first token (TTFT) measures how long a streaming request waited before its first output. Streaming duration measures the time spent sending the answer. Finalization covers work after the stream ended, when reported.

Good to know: A slow first response and a long answer are different problems. Compare the same model and similar requests, then check TTFT and stream timing. For pauses while a person answers, inspect interrupt timing.

Compare a broader period

Use Workflows to inspect timing across recorded runs in the selected period. Median (p50) is the middle observation; p95 describes the value that 95 out of 100 observations fall at or below. Check the sample count before treating a percentile as representative.

Next: Metric definitions or token usage and cost.