Statistics page

The Statistics page aggregates request usage, breaks it down by model/key/group dimensions, and supports time-series display. Data is persisted to the database at 1-minute granularity.

Entry: Console → Statistics (visible to admin and monitor).

Summary cards

Seven cards at the top, reflecting totals for the selected period:

Card	Meaning
Total requests	All requests in the period
Successful requests	2xx
Error requests	4xx + 5xx
Average latency	milliseconds
Input Token	Input tokens counted by the upstream
Output Token	Output tokens counted by the upstream
Cache read Token	Tokens hit in the prompt cache

Dimension switching

group_by switches the breakdown dimension:

Dimension	Main table first column
model	By model
access_key	By access key (expandable)
access_key_group	By key group

Remaining columns of the main table: request count, success count, error count, success rate, average latency, input tokens, output tokens, cache read tokens.

Time range

• Presets: 24h (default) / 7d / 30d
• Custom: pick start and end times with the DateRangePicker, supporting relative expressions

Time-series chart

Granularity is chosen automatically based on the selected period (targeting about 120–720 data points):

Time span	Granularity
≤ 24h	5 minutes
≤ 2d	15 minutes
≤ 7d	1 hour
≤ 30d	4 hours
≤ 90d	12 hours
> 90d	1 day

Shows curves of token usage (input+output by model) and request count (total/success/error).

Export Excel

The export entry is on the billing page (Console → Settings → Billing → Monthly costs): pick a month and click export to get an Excel with 6 sheets (model / access key / key group / access key × model distribution / day / week time dimensions), with costs split into columns by currency. See Export billing Excel and log troubleshooting for details.

Notes

• count_tokens requests (the Anthropic token-counting endpoint) are not counted in statistics, because they do not actually call the upstream and produce no cost.
• Persisted statistics are stored in 1-minute buckets; the current minute bucket is lost if the gateway crashes, while completed minute buckets are retained.
• Statistics and billing record the final mapped model name. If an access key / key group has an identity-scoped model mapping configured (e.g. claude-opus-5 → qwen3.7-max), the Statistics page, Excel export, and bills will only show rows with the target name qwen3.7-max and will not show rows with the source name claude-opus-5 — mapping is fully transparent to statistics and billing. If you need to attribute by the client-side source name, record it on the client side yourself. See Configure identity-scoped model mapping for details.

FAQ

Q: Request counts in statistics don't match the logs? Statistics are aggregated at 1-minute granularity, while logs are per-request records. Within a short time window there may be up to 1 minute of flush delay. count_tokens requests appear in logs but are not counted in statistics.

Q: Token counts are high? Confirm whether count_tokens was also counted — it is not counted in statistics. If still high, check whether an upstream returned an inflated usage field.

Q: I configured model mapping, why can't I see the model name the client requested in statistics? Because statistics and billing use the final mapped model name. When the client requests claude-opus-5 and it is mapped to qwen3.7-max, statistics only record qwen3.7-max. This is expected behavior; see Configure identity-scoped model mapping for details.

Next: Overview dashboard for a global view; Export billing Excel and log troubleshooting for the export operation; Log viewer for per-request details.