Does Quasar know which model an agent used, and how many tokens?
Yes, from the agent's own client, not from its word:
- Set up once per client. With an ingest key from the Profile, the client sends Quasar its own telemetry: each turn, and each call to a model with its model and tokens. It never sends the text of a prompt or of an answer. Measured clients:
- Claude Code and the Claude desktop app;
- Codex CLI;
- opencode 2;
- pi, with its pi-otel extension.
- On the work log. A confirmed turn shows the tokens of its own share, model by model, «measured on the turn».
- On a plan's step (since 2026.11.10). Each step shows the model of the turn that closed it.
- Without telemetry, Quasar does not know, and says so: a step shows «model not reported». Since 2026.11.13 a person whose agent works through Quasar while its client's telemetry does not arrive is told so by the notification Telemetry not arriving, and a work log says when the provider reported no cache reads.
Sessions agents declared before 2026.11.10 keep the comparison between the usage they declared and the usage their client reported.