Cost findings

Cost findings shows dollar figures from your first traces, before any outcome is labeled. Every item is a candidate to review, not a conclusion. A run with no outcome recorded is unknown, not failed. Morse does not say it was a bad run, and it never claims to have saved you money.

Open it from the Cost findings tab under Cost & Budget in the sidebar, or from the First cost findings card on Home.

The word “finding” on its own means a case your team decides on — see Follow-ups and decisions.

The default window is the last 7 days.

How the page is grouped

Findings sit in separate groups by how the number was produced. Morse never adds numbers from different groups into one total, because a measured dollar and an estimate are not the same kind of figure.

Measured in your traces

These are dollars your spans already recorded, priced the same way as on Cost Intelligence. Within the group they are ranked by dollars.

• Repeated tool calls and retry loops. These are spans where the same tool ran again inside one trace:

  • with identical arguments, or
  • as three or more consecutive attempts after failures.

  The figure is the priced cost of the repeats only, not the whole trace. A repeat with no price is counted but not given a dollar amount. Tool-call fingerprints are recorded from ingest onward and older traces are not backfilled, so this check starts with new traffic. See Efficiency.

• Input spend by context segment. This shows what share of an agent’s billed input spend went to each part of the prompt, such as the system prompt, tool definitions or history. It covers the three agents with the highest spend. Each model call’s input tokens are split across segments by estimated size, then priced at the rate that applied (cache read, cache write or standard input). The figure is what that segment cost, not what a change would return; whether quality holds after a change is only known once you test it. See Context & Memory.

• Where spend went. These are the agents and traces with the highest priced spend in the window. This item is always shown once there is one priced trace. It is a plain measurement and makes no claim about whether the spend was good or bad.

Estimates at list price

• Prompt caching. Morse looks for an agent that repeats a long, identical prompt prefix on Anthropic models. The agent qualifies when:

  • it made at least 20 calls,
  • at least 60% of those calls share the prefix, and
  • at most 5% of calls were served from cache.

  The estimate uses provider list prices and the cache time-to-live (5 minutes or 1 hour) that gives the better result. It counts a cache write for the first call and for every gap longer than that time-to-live. It is shown per month. Your contract price may differ. See Context & Memory.

Worth a look

• Larger-than-usual context. These are traces whose context was at least 4× the agent’s usual size, measured against a 14-day baseline that needs at least 5 earlier traces. Only the count and the trace links are shown. There is no dollar figure, because the trace’s whole cost is not the cost of the extra context. Morse checks the 10 agents with the highest spend in the window. A trace already listed under repeated tool calls is not listed again.

Outcome coverage

This group shows the share of spend in the window that has no confirmed outcome yet. It is there to tell you what unlocks verdicts. It does not mean those runs failed, and the page never shows cost per success. To fill it, record outcomes with set_outcome (see the SDK docs).

Checks that are still forming

If a check cannot run yet, it appears in a single line that lists what it needs and the one action that fills it. Examples:

• no tool-call fingerprints yet
• fewer than 5 baseline traces
• no Anthropic calls

A check that cannot run is never shown as a $0 finding.

For checks still forming, Evidence ETA shows the missing evidence, an estimate range when available, the next action, and labeled/total coverage.

Download PDF

Download PDF exports the same findings shown on screen. The PDF includes each item’s method, assumptions, sample size and trace IDs, and ends with a methods section. Only signed-in members of your organization can download it. There are no public share links.