Telemetry

Telemetry is the raw-data view. Where the other pages interpret your work, this one just shows the numbers — cost, tokens, requests, errors, and session durations — for your own sessions, organized into tabs. Come here when you want to check a figure directly or dig into a spike you noticed elsewhere.

The headline tiles

Four tiles sit above the tabs, covering the range you pick:

• Total Cost — the API-equivalent cost of your usage. When a model can’t be priced it shows —, never a misleading $0.00.
• Sessions — how many sessions you ran.
• Avg Cost / Session — total cost divided by sessions.
• API Requests — how many requests you made.

The five tabs

There are five tabs — Cost · Tokens · Requests · Errors · Duration — and each tab’s charts let you set the granularity (hour · day · week · month) per chart.

Cost

• Spend Over Time — spending across the period, on a time axis.
• Cost Distribution — where cost concentrates.
• Model Breakdown — a table of cost, requests, average per request, and share by model. Only priced models are counted.

Tokens

Six token tiles come first:

• Total Tokens · All usage — total tokens used in the period.
• Fresh Input — input tokens the model hadn’t seen before.
• Output Tokens — tokens the model generated.
• Cache reuse — context served from cache, which costs less.
• Cache write — tokens newly written to cache.
• Avg Tokens / Request — the average tokens exchanged per request.

Below them come three charts:

• Token Usage Over Time — an area chart of token usage across the period, split into input, output, and cache.
• Token Distribution — a donut of each model’s share of the total tokens.
• Model Token Breakdown — a per-model table of input, output, cache reuse, and cache write tokens.

Requests

• Requests Over Time and Sessions Over Time — request and session counts per period.
• Activity Heatmap — a day-by-hour grid of when your requests cluster.

Errors

• Errors Over Time — error counts per period; a downward trend means cleaner runs.
• Errors by Status Code and Errors by Model — errors split by status code and by model.
• A Recent Errors list and Retry Patterns — the latest errors and how often the same request was retried.

Duration

• Avg Session Duration, Median Session Duration, and Avg LLM Time tiles.
• Avg Session Duration Over Time and a Duration Distribution.
• LLM vs Idle Time — how session time splits between the model working and waiting.
• Longest Sessions — your longest-running sessions.

Reading it

Pick a date range and, per chart, a granularity (hour / day / week / month). Charts leave out unpriced or zero-value models rather than drawing a misleading “0% of spend.”