# Run Stats card

> One agent's run in exact numbers — prompts, model requests, tokens by kind, elapsed and working time. The official card for benchmarks.

Source: https://docs.neurosquad.ai/en/cards/run-stats

**Run Stats** measures one agent's run and shows it in big, exact numbers you can screenshot or
paste into a report. It is the official NeuroSquad card for benchmarks: give Claude Code, Codex
and OpenCode the same prompt on the same model, put a Run Stats card next to each, and compare.

It is a verified [custom card](https://docs.neurosquad.ai/en/card-sdk/verified-cards) by NeuroSquad. In the add menu (**+** on the canvas)
pick **Custom card…**, install *Run Stats* from the **Verified** tab, add it to the canvas and draw an
arrow between the card and an agent. Requires NeuroSquad 0.1.254 or newer.

## What it shows

| Number | Meaning |
| --- | --- |
| Prompts | Turns the agent was given. Answering a permission prompt continues the same turn. |
| Model requests | API round-trips the agent made, from the harness's own log. |
| Input | Uncached input (prompt) tokens. |
| Output | Output tokens, reasoning included (the reasoning part is shown under it). |
| Cache read / Cache write | Input tokens read from / written to the provider's cache. |
| Total tokens | Input + output + cache read + cache write. |
| Elapsed | From the first prompt to the last finish (ticking while the agent works). |
| Working time | Time the agent was actually working — waiting on you is excluded. |
| Cost | At the list price, or what the harness itself recorded. Unknown model — *no price*, never $0. |

Token numbers are exact integers, the same ones the [Usage](https://docs.neurosquad.ai/en/usage) section uses. When a harness
or the server behind it does not report a number — for example a local model server that sends no
cache counts — the card says **not reported** instead of showing 0. Prompts, elapsed and working
time work for every agent; Amp and Cursor keep no usage log NeuroSquad can read, so for them only
those are measured.

## Measuring a run

- When you connect an agent, the card counts everything that agent did so far — for a fresh agent
that is exactly the run.
- **Start measuring now** (later **New run**) starts a clean measurement: the run is the next
prompt.
- **Freeze at finish** (on by default) pins the numbers when the agent finishes, so they stay put
for the screenshot even if you keep talking to the agent. **Unfreeze** to keep counting.
- **MD** and **JSON** copy the results; **Send to note** (card menu) appends a Markdown table to a
note connected by an arrow.
- Zoomed out, the card's tile shows total tokens and elapsed time.

**Good for:** comparing CLIs or models on the same task, and answering "what did that run cost".