Skip to content
Enso

Stats

What it shows

Section titled “What it shows”

The Stats tab analyses one thread: what it cost, how the cache behaved, which parts of the conversation were expensive, and which tools did the most. Every figure comes from pi’s own session file, so a thread from last month gives the same answer as one from a minute ago.

How to open it

Section titled “How to open it”
  • Open a thread from the rail, then choose the Stats tab at the top of the thread.
  • A point on a chart, a call or a prompt in a list opens that spot in the chat.

What it looks like

Section titled “What it looks like”

The Stats tab

  1. failures, and under it the refusals, counted apart: 2 refused by a guard — not failures.
  2. Tokens: each turn’s cache read, cache write, cache miss and output, as a stacked bar.
  3. Cost per turn, with the median beside the title.
  4. Costliest calls: the tool calls ranked by input, output, time or failures.

Further down: Cumulative usage by tool, Distribution, Costliest prompts and Latency, then the thread’s Session eval.

What to look for

Section titled “What to look for”

The thread above cost $0.2128. Three things stand out.

The first turn after the switch to claude-opus-4-1 is one solid cache-write bar in Tokens, and the peak of Cost per turn: the new model had no cache, so the whole prefix was written again, at that model’s price. Nearly every other turn is almost all cache read. If you see a lone cache-write bar in a thread of your own, look for a model switch or a compaction just before it. Context Events names which.

The line under the metrics, $0.0286 not attributed to a turn on this branch, is the compaction. Its summary was a model call of its own, but not a turn of the conversation, so the tab names the cost instead of hiding it in a turn.

failures reads 1 · 12.5%, not counting the refusals. pi records a guard’s refusal as an error, just like a real failure. Enso keeps a record of each refusal, so the tab counts the failed cat as a failure and the refused calls apart. A refusal is the guard working, not the tool breaking.

What it doesn’t do

Section titled “What it doesn’t do”
  • It covers the current branch only. Cost spent on a branch you forked away from, or on a compaction, is named as not attributed to a turn, rather than spread over the ones you see.
  • Times come from the log, which keeps 14 days. An older thread still has every token and cost figure, but its Latency says no day file covers it.
  • Tool figures are in characters, not tokens. A tool’s cost is the tokens its result adds to later requests, which the Context tab shows.

Go deeper

Section titled “Go deeper”