Fields
Status + cost
Tokens
Input + output
Export
CSV
One row per call with status, tokens, cost, latency, model, and provider.
New capabilities
Filter by status, provider, model, key, and time window to land on the exact calls behind an incident instead of grepping app logs across services.
Status
424
Provider dependency failed; the row links to the model and key that hit it.
Latency
1.8s
Response time for the call, so slow successes are separated from failures.
Cost
$0.041
Estimated request-level spend from token count and model pricing.
Each row carries input and output token counts, estimated cost, and the key that made the call, so a spend number always traces back to real requests.
Organization log views can show metadata — status, tokens, cost, latency — while hiding prompt and response bodies, so on-call engineers can debug without reading sensitive user data.
Who Concentrate is designed for
A request log is one row per model call: status, timestamp, model, provider, and the key that made it, plus the token breakdown (input, output, cached, and total), estimated cost, and duration. Status separates clean successes from failures like a 424 when a provider dependency errors out. Every spend number and usage chart traces back to these rows, and it's where on-call engineers go first when an AI feature breaks.
Every request through the gateway is recorded with status, time, model, provider, key, token counts, cost, and latency in a single filterable view.
Because traffic to every provider passes through the gateway, you filter one log instead of stitching together separate OpenAI, Anthropic, and Google dashboards.
Org-level views can expose metadata while hiding prompt and response bodies, so debugging doesn't require reading sensitive content. Pair it with data redaction for stronger controls.
Feature basics
Spend tracking and usage analytics aggregate these rows, so the totals always trace back to the calls underneath them.