Shift Note: Trace the Requests Behind the Spend

A spending total tells you that money left the account. A traffic log helps you find which requests caused it. For someone new to model APIs, the useful starting point is simple: find the expensive calls, then look at what the application asked them to do.

One costly request might be doing valuable work. Hundreds of small requests might be repeating the same mistake. You need to inspect both the price of each call and the pattern around it.

Start with a small window of traffic

Choose a period you can connect to actual activity, such as the hour when you tested a coding agent or processed a batch of documents. A full month of traffic can hide the relationship between an action and its requests.

Depending on your provider or logging setup, useful fields may include the timestamp, model, input tokens, output tokens, status, request identifier, and charge. If charges are available, sort by them from highest to lowest. If they are missing, token counts and the applicable model prices can help you estimate cost.

Tokens are pieces of text that the model processes or generates. Input tokens include material sent to the model. Output tokens cover generated content, although providers may report additional usage categories. Check the billing definitions before treating a token count as a complete cost calculation.

Read the largest charges as individual cases

Open a few of the most expensive entries and connect them to the task that triggered them. Avoid assuming that the model name explains everything. Cost also depends on the amount of work sent, the amount generated, and the applicable pricing rules.

• Check whether a large input came from a long conversation, a pasted document, retrieved passages, or tool results.
• Check whether a large output was useful or whether the application requested a lengthy answer for a small task.
• Check whether the chosen model matched the difficulty and quality requirements of the task.
• Check whether caching, reasoning usage, or other separately priced categories affected the charge, where those details are available.

For example, a short classification task sent to Claude Opus 5.5 is worth reviewing for a possible move to Claude Haiku 4.5. Test representative inputs first: a cheaper call only helps if it still produces an acceptable result.

Look for repeated calls below the expensive ones

Sorting by individual charge catches large requests. It can miss a loop that makes many modest calls. Group traffic by a field that connects it to the work, such as a session, application, job, or route, if your logs provide one.

Look for clusters with similar timestamps, repeated input sizes, or repeated failures. A single user action can trigger several model calls when an agent uses tools, checks its work, or retries. Some of that activity is expected. An unexpected jump deserves investigation.

A failed request does not automatically mean it was free, and similar entries do not automatically prove duplicate billing. Compare request identifiers, application logs, retry settings, and the provider's billing rules before deciding what happened.

Make one change and check the next run

Choose the clearest cause and make a targeted adjustment. Remove irrelevant material from prompts, limit answer length where appropriate, fix unnecessary retries, or test a different model for a narrow task. Avoid changing everything at once; that makes the next result harder to explain.

Compare the same kind of work before and after the change. Check total cost, request count, and whether the output still meets your needs. A lower charge per request can be misleading if the application now needs more attempts.

Keep a repeatable review across model families

Relay365 is worth considering if your workflow uses several model families and you want pay-as-you-go access through one relay service. It works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients. Create a Relay365 account when you are ready to test it with a small, representative workload.

Keep a short shift note recording the costly pattern, the change you made, and the result. That gives the next traffic review a concrete starting point.