YOUR MODELS. ONE CLEAR VIEW.
Tokens tell a story.
See the whole picture.
Loading last successful collection…
Connecting…
Models · all
Daily tokens and cost
Colored bars: tokens by model, left axis. Black line: daily total estimated USD, right axis.
No usage matches these filters.
Monthly highest-cost model
Highest recorded estimated cost within the selected dates and models. Cost share is relative to the period total; zero-cost models are not cost leaders.
Per-response speed analysis
TPS = output tokens / response seconds, including first-token wait; averages and percentiles use individual responses, not session duration. First-token latency uses only recorded first-token timing, not total response time. Missing or zero timing stays unavailable. Only successful responses with output are included; historical rollups have no response timing.
≈ includes native-log response-window estimates, not streaming-only generation speed. Disable the estimates above to use only recorded request timings.
Period comparison
Last 7 and 30 days include today, in dashboard host time. All three rows also respect the date and model filters above.
| Group | Responses | TPS samples | Mean TPS | Median TPS | Max TPS | First-token samples | Mean first token (s) | Median first token (s) | P95 first token (s) |
|---|
Daily average response speed
Weekday × hour of day
Averages over individual responses at each weekday and hour in dashboard host time. Gray tiles have no available timing for the selected metric.
Hover or focus a tile for values and sample counts.
By model
| Group | Responses | TPS samples | Mean TPS | Median TPS | Max TPS | First-token samples | Mean first token (s) | Median first token (s) | P95 first token (s) |
|---|
By weekday
| Group | Responses | TPS samples | Mean TPS | Median TPS | Max TPS | First-token samples | Mean first token (s) | Median first token (s) | P95 first token (s) |
|---|
By hour of day
| Group | Responses | TPS samples | Mean TPS | Median TPS | Max TPS | First-token samples | Mean first token (s) | Median first token (s) | P95 first token (s) |
|---|
Per-project usage
Projects with the same name and path are combined across machines and apps. Recognized cloud-drive roots use cloud-relative paths; different folders or cloud services remain separate. Date and model filters apply. Missing or conflicting metadata stays unassigned.
TPS uses the same response-time calculation and native-log estimate switch as per-session analysis; untimed responses are excluded.
| Project | Machine / app | Models | Fresh input | Cache read | Cache write | Output | Total tokens | Requests | Sessions | Est. USD | Avg TPS | Max TPS | Timed responses |
|---|
No project detail matches these filters. Refresh collection to load saved project metadata.
Project detail
By date
| Date | Fresh input | Cache read | Cache write | Output | Total tokens | Requests | Sessions | Est. USD | Avg TPS | Max TPS | Timed responses |
|---|
By model
| Model | Fresh input | Cache read | Cache write | Output | Total tokens | Requests | Sessions | Est. USD | Avg TPS | Max TPS | Timed responses |
|---|
Sessions
| Conversation | Machine / app | Total tokens | Requests | Est. USD | Avg TPS | Max TPS | Timed responses |
|---|
Per-session token usage
Only usage inside the selected dates and models. Conversation names are saved Codex / Claude Code titles. Identically named sessions remain separate; some source records have no matching saved title.
≈ includes Codex estimates: output tokens divided by the logged user/tool-input-to-generated-output window, not the whole turn. Completed tool execution and gaps between turns are excluded; client scheduling and time to first token can remain. Only unique exact session, usage timestamp and token-count matches are used. Missing boundaries stay unavailable. This switch affects TPS only and is saved in this browser.
TPS = output tokens / recorded response seconds (duration_ms, otherwise latency_ms), including time to first token. Average is the arithmetic mean of successful timed response rates; maximum is the fastest response rate, not peak streaming speed. Session idle time and input/cache tokens are not used. Untimed or zero-output responses are excluded; — means unavailable. Date and model filters apply.
Temporal Session Token Usage
Smooth colors interpolate between adjacent recorded time buckets, not additional measured activity. Gaps stay blank. Hover or focus for exact bucket totals. Includes input, cache and output tokens; historical rollups without session detail are excluded.
Hover or focus a colored cell for its title, date and exact token count.
| Conversation | Selected activity | Machine / app | Models | Fresh input | Cache read | Cache write | Output | Total tokens | Requests | Est. USD | Avg TPS | Max TPS | Timed responses |
|---|
No identified sessions match these filters.
Session detail
Token components
| Component | Tokens | Share of session |
|---|
By date
| Date | Models | Fresh input | Cache read | Cache write | Output | Total | Requests | Est. USD | Avg TPS | Max TPS | Timed responses |
|---|
By model
| Model | Fresh input | Cache read | Cache write | Output | Total | Requests | Est. USD | Avg TPS | Max TPS | Timed responses |
|---|
Detail follows the active date and model filters. Cost is the recorded total estimate; the source does not attribute it separately to input, cache, and output.
Sources, freshness & definitions
Recorded costs are estimates, not invoices or subscription fees. Zero cost can mean missing pricing. Timestamped requests are aligned to the dashboard host’s local time. Historical rollups keep their source-local date. Incomplete months use available days.