Skip to content
Free$1 of credit for every new organization
All articles

Product

Reading the Logs page: model, provider, tokens, cost, latency

Every request you send shows up in the dashboard Logs. What each column means, and the questions it can answer about cost, speed and failures.

· 3 min read · Usefully team


The Logs page in the dashboard lists every request your organization sends through Usefully, chat and search alike. It's the fastest way to answer the questions that come up after launch: why did spend jump yesterday, which feature is slow, did that outage affect us? Here's what's in it and how to use it.

What each entry shows

  • Model: the model you asked for, such as anthropic/claude-sonnet-5-5.
  • Provider: who actually served it. If a request failed over, this is the provider that succeeded.
  • Tokens: input and output tokens, as reported by the provider.
  • Cost: what was charged from your balance for this request, at the provider's list price.
  • Latency: how long the request took end to end.
  • Key: which API key sent it, by name.
  • Status: success, failed, or partial (a stream that was cut short).

Logs are metadata only. Prompt and response bodies are not stored, so you won't find the text of a request here. If you need to correlate a log entry with your own records, log the response id in your app.

Four questions come up again and again, and Logs answer each of them.

"Why did spend go up?"

Start with the spend charts: spend by day shows when, spend by key shows which service, spend by model shows what. Then open Logs for that day and look for the most expensive requests from that key. The usual culprits are a prompt that grew (look at input tokens), a loop that retried a successful request, or traffic moved to a more expensive model.

"Why is this feature slow?"

Find requests from the key or model the feature uses and look at latency next to output tokens. Latency that tracks output length is generation time: shorten answers, lower max_tokens, stream, or try a faster model. High latency on short answers points elsewhere, such as very long inputs. See reducing latency.

"Did the provider outage affect us?"

Look at the provider column around the time of the incident. If requests for a model were served by a different provider than usual, failover did its job and your users saw a successful response. Failed requests show their error type and aren't charged.

"Is this key still in use?"

Before revoking an old key, check Logs and spend by key for it. If nothing has arrived for a day, it's safe to revoke. That's the core of rotating keys without downtime.

Habits worth building

  1. 1Name keys by service and environment (checkout-bot-prod, evals-staging). Logs and spend charts become self-explanatory.
  2. 2Check spend by model weekly in the first month after a launch. Surprises are cheapest when caught early.
  3. 3Set a low-balance alert so the first sign of trouble isn't 402 insufficient_credits in production.
  4. 4Record `request_id` from errors in your own logs; support can find the request from it.

Logs also make a good first stop when a teammate reports "the AI is broken": the status and error type usually tell you within a minute whether it was a provider issue, a limit, or a bad request from your code.

Logs are available to every member of an organization, including viewers. Open your dashboard to see them, or create an account to get started with $1 of free credit.

More articles

Get started

Your first request is five minutes away.

  • $1 of free credit to start
  • No card, no subscription
  • Keep the SDK you already use
  • Real people at support@usefully.info
Reading the Logs page: model, provider, tokens, cost, latency · Usefully