| Token Usage | Shows how token usage changes over time. | Use this to locate usage peaks. If a peak looks abnormal, compare it with application logs and request records in Usage Detail. |
| API Requests | Shows how the number of API requests changes over time. | Use this to understand traffic changes. If requests increase but token usage stays mostly flat, short requests or probe-style requests may have increased. |
| RPM | Shows requests per minute over time. | If RPM rises sharply, request volume or concurrent usage has increased. Compare it with the RPM limits on your API keys. |
| TPM | Shows tokens per minute over time. | If TPM rises sharply, overall token throughput has increased. Check Token Distribution to see whether input, output, or reasoning tokens are driving the change. |
| Token Distribution | Shows the distribution of token types, such as input, cache read, cache write, reasoning, and output tokens. | Use this to analyze cost composition. If output or reasoning tokens take a large share, review generation length, reasoning model usage, and parameter settings. |
| Prompt Bucket Distribution | Groups requests by prompt length bucket and shows the corresponding TTFT and E2E latency percentiles. | If long prompts account for a large share of requests, consider compressing context, splitting tasks, or using Prompt Caching to reduce repeated context processing. |
| TTFT | Shows the time from request start to the first output token. | If TTFT increases, model API callers may notice that responses start more slowly. Analyze this together with model choice, prompt length, streaming behavior, and traffic peaks. |
| E2E Latency | Shows the total time from request start to response completion. | If E2E latency increases, it is often related to output length, model type, context length, or concurrent load. Compare it with output token counts and request records in Usage Detail. |
