Analytics

The analytics endpoints read the gateway's request telemetry, which is a separate store from the control plane's own records. Every call is scoped to the caller's organisation from the session. If the telemetry store is unreachable they answer 503 {"error": "analytics backend unavailable"} while every other endpoint keeps working.

Common filters

Blank values equal omitted ones. Averages are null, not 0, over a window with no matching requests.

GET /analytics/summary

{
  "time_range": "24h",
  "total_requests": 1420,
  "avg_e2e_latency_ms": 812.4,
  "avg_ttft_ms": 143.2,
  "avg_tokens_per_sec": 61.3,
  "total_token_usage": 9412338,
  "total_cost_pkr": 61.05
}

time_range echoes the window applied; label panels from the response, not the request. total_cost_pkr (plain PKR) is present only when the query names a key_id; on an organisation-wide summary the field is absent. Tokens per second is derived at query time from tokens generated over generation time.

GET /analytics/timeseries

Parameters: metric (volume, latency, ttft, tok_s, token_usage), interval (bucket width, 30s, 5m, 1h, 1d), group_by (project, key, model), plus the common filters.

{
  "metric": "latency",
  "interval": "1h",
  "group_by": "model",
  "time_range": "24h",
  "data": [
    {"bucket": "2026-08-25T19:00:00", "group": "IntelligentInference/gpt-oss-120b(Global)", "value": 1200.0}
  ]
}

One row per bucket and group, a flat shape you can pivot however your charting library wants. group is the raw project id, key id or route id; resolve names with /projects and /keys. Bucket timestamps carry no timezone and are UTC. The endpoint does not stop you pairing interval=30s with time_range=30d; pick a sane pairing.

GET /analytics/requests

The request log: one row per request, newest first, paged.

{
  "total": 1420,
  "limit": 50,
  "offset": 0,
  "requests": [
    {
      "req_id": "k2i-1787578888306946000-0001",
      "timestamp": "2026-08-24T12:34:56.789Z",
      "model": "IntelligentInference/gpt-oss-120b(Global)",
      "status": 200,
      "stream": true,
      "key": {"id": "e02dcba3-…", "name": "checkout-summariser-prod"},
      "project": {"id": "3dbb877d-…", "name": "Assistant"},
      "metrics": {
        "total_ms": 420,
        "ttft_ms": 110,
        "tpot_ms": 6.8,
        "pre_dispatch_ms": 5,
        "prompt_tokens": 45,
        "tokens_generated": 40,
        "cache_creation_input_tokens": null,
        "cache_read_input_tokens": null,
        "total_tokens": 85,
        "tokens_estimated": false
      },
      "financials": {"cost_pkr": 0.04512, "is_byok": 0},
      "status_detail": {"termination_reason": "success", "upstream_error": null},
      "upstream": {"model": "openai/gpt-oss-120b"},
      "user_agent": "OpenAI/Python 1.51.0",
      "metadata": {"user_tag": "emp-4092", "feature_tag": "ticket_summary"}
    }
  ]
}

Nulls are meaningful: key or project is null when the request had no attribution at all (as opposed to an object whose name is null because the key has since been deleted); metrics.ttft_ms and tpot_ms are null for non-streaming requests and for a stream whose first-token event has not flushed yet; metrics.prompt_tokens and the whole upstream block are null on rows served before 2026-08-24, when the gateway began recording them.

Prompt caching, billing mode and client

Both cache counts are null on i2's own global routes, which have no prompt cache. null there means the route does not cache, which is not the same claim as 0, and a dashboard or report that treats them alike will show a route with no cache as a route that never hits one.

is_byok is the explanation for a near-zero cost_pkr. On a BYOK request the provider bills you directly and i2 settles only its own share, so a row can legitimately cost a fraction of a paisa. Read the cost against this flag, not on its own.