Experiment lifecycle
Define variants
Compare a system prompt, model, technique configuration, or a controlled mix of all three.
Loading
LoadingCheaper Inference / Optimization / Experiments
The experiment keeps assignment and measurement together, so every result can be traced back to the requests that produced it.
Compare a system prompt, model, technique configuration, or a controlled mix of all three.
Use weighted random assignment per request or sticky assignment per end user.
Read cost, input and output tokens, request volume, savings, and quality feedback together.
End the experiment and make the winning variant the new production baseline without a redeploy.
Evidence / 02
Cheaper is useful only when the result still works. Cost and token use stay beside workspace feedback, so a low bill cannot hide a worse answer.
Cost
$0.011
Average billed cost
Tokens
684 → 312
Input to output
Feedback
4.8
Average quality score