pixel

Savings estimate

What Pixel could spare your team's agents each month on large-file reads, on your own numbers. An estimate, computed in this page; nothing leaves it.

Updated

Everyone whose agent works in your repositories.

Agent sessions started, not prompts: a session usually spans several turns.

Whole reads of a large file to learn what it holds. Short files and targeted line ranges do not count.

Bytes divided by four. 10,000 is about 1,000 lines at the roughly ten tokens a line of the well-known files, which run 13,995 to 63,908 tokens each.

Your model's list price for uncached input tokens. The default is a round placeholder, not a quote.

Five days a week, less a public holiday.

Estimate, per month

$119 of input tokens not billed

range $100 to $122

39.7M tokens your agents do not read

range 33.5M to 40.8M

At the 8 well-known files' rate: 94.5% median, 79.7 to 97.2% (the files).

This rate is for reading one large file, not for a whole session, which also spends tokens on prompts, edits, tests and answers. On a whole agent task we measured −30% API cost and −38% tokens read, the median of 11 runs per side (the runs).

Show the math
  1. Reads a month = 10 developers × 4 sessions × 5 reads × 21 days = 4,200
  2. Tokens read whole = 4,200 × 10,000 tokens = 42,000,000
  3. Tokens not read = 42,000,000 × 94.45% (median) = 39,669,000; at 79.7%: 33,474,000, at 97.2%: 40,824,000
  4. Dollars = tokens not read ÷ 1,000,000 × $3 = $119.01; from $100.42 to $122.47

The three rates are the saving on each kept file of the well-known files table: the whole file against pixel list-signatures on it, UTF-8 bytes divided by four on both sides, Pixel's answer already subtracted. The median of an even count is the mean of the two middle files.

What it leaves out: a file read stays in the context and is sent again on every later turn (usually from the cache, billed lower), so a real session pays for it more than once; output tokens do not change; and the files measured run 13,995 to 63,908 tokens, so much smaller files may save less.

Share this estimate

The link carries your six values in its fragment (#d=…&s=…), which the browser keeps to itself: opening it recomputes the same estimate, and nothing is sent to us or stored.

https://pixel-cli.dev/savings/#d=10&s=4&r=5&t=10000&p=3&w=21

A README badge with the median, linking back to these values. The image is a static shields.io badge, drawn from the text of its URL.

[![Pixel estimate: 39.7M tokens a month not read](https://img.shields.io/badge/Pixel%20estimate-39.7M%20tokens%20a%20month%20not%20read-22c55e)](https://pixel-cli.dev/savings/#d=10&s=4&r=5&t=10000&p=3&w=21)

To measure instead of estimate, run pixel token-savings in your repository after a few sessions: it reads Pixel's local action log. Measure it on your own code.