We find ourselves currently struggling with using Claude Sonnet confidently because under heavy load the response delay from Anthropic Servers gets quite extreme.

We measure the perceived latency through manual testing.

Current Information: In the LLM provider choice options Hume applies labels like (fast or faster) to some models, but this does not reflect actual real world performance fluctuations.

It would be a great help to have a dashboard of realtime latency measurements for each of the LLM provider options.

The values could be lazy refreshed on new calls occasionally to reach sufficiently recent measures at all times.