Rate card · your team, today
Human focus is the mostvaluable resourcein the building.
Reading is an engineer's input rate. Writing is the output rate. Price that attention the way you price an API and you get a number worth having in front of you before you route another document through six people.
is what an engineer's attention is worth over a page of reading, measured against Claude Fable 5. On output it is 750×. That is not an argument for reading less. It is an argument for being deliberate about what reaches a person at all.
Step one
Calibrate the two rates
The defaults are population averages and your team will differ. Both tests below run in under a minute and write their results straight into the rate card.
Reading speed
238 wpm by default. Brysbaert (2019), a meta-analysis of 190 studies covering 18,573 people, silent reading of English non-fiction. Start the clock, read the passage the way you would read anything, stop it at the last word. Capped at 900 wpm, which is already past anyone reading rather than skimming.
The passage appears when you start the clock. Reading it early only costs you twice.
Typing speed
40 wpm by default, rounded up from the 36.2 wpm measured by Palin et al. (2019) across 37,370 volunteers. Rounding up makes you look cheaper, which is the direction an honest assumption should err. Capped at 400 wpm.
Step two
Itemise something you already paid for
Pick a preset or paste the real thing: the diff you reviewed, the RFC you commented on. Words are counted in this tab and the text never leaves it.
The reasoning a model does between reading and replying, billed as output — a share of everything the job processes. It moves the receipt below and the passes further down. Pure reading does none, so a read-only receipt does not change.
An average pull request
Small on its own. This is the unit that repeats every week, which is the reason to price it at all. Code is read slower than prose, so this understates the reading. Paste a real diff for the real number.
tokens ≈ words × 4/3 · 238 wpm reading, Brysbaert 2019
computed in your browser, sent nowhere
Step three
Getting the content right, priced two ways
One draft, worked over until it is right. Passes through a model on one side, passes through people on the other, both priced off the same rate card. Change any number and both sides move.
Standing context the model re-reads each pass. A person keeps this in their head; it caches.
5 to 10 is the usual band on a draft that has to be right.
6 model passes
$2.79
Claude Fable 5 list, $10.00 in and $50.00 out per 1M, checked 19 August 2026. Each pass reads the draft plus the code and standing context above, and writes it back plus its thinking — the 40% share set with the receipt above, billed as output. Standing context caches at a tenth after the first pass ($0.50 up front).
2 human passes
$45.17
Priced off the rate card above: $120 an hour, 238 wpm reading, 40 wpm writing. 23 min of somebody's week.
Same draft, two ledgers: $2.79 of model passes against $45.17 of human ones. The cheap one is there to protect the expensive one: run the model passes so the human passes land on the part of the work that actually needs a person. That is a price on passes and nothing else. It says nothing about which pass catches the thing that matters, and this page does not have that number.
What this is
Three things, in the order they matter
There are already dashboards that count how many pull requests closed. This one is built on a different unit, a different scope, and the opposite conclusion.
Priced in tokens
Engineering metrics count merged units and cycle time. Tokens are the unit your finance team already understands, because they are on the invoice from the model vendor. Reading is a price per million in. Writing is a price per million out. Both sit on the same table as the model's.
One thing at a time
Not a quarterly productivity total handed back as headcount. One pull request, itemised, with the machine price beside it. $33.67 for a single review is a number you can hold against the twelve of them queued for Thursday.
Expensive is the good news
Every other version of this ends in a lecture about focus, and nobody has ever changed a calendar because of one. An expensive hour is not waste. It is the valuation. Nothing here says review less; it says know what a seat costs before you put one in the loop.
Show the working
Every number on this page, and where it came from
The arithmetic
Time is what you actually spend. Dollars come out of time, and tokens describe how much content that time bought.
minutes = words ÷ wpmdollars = minutes ÷ 60 × your hourly ratetokens ≈ words × 4/3$ per 1M tokens = hourly ÷ (wpm × 60 × 4/3) × 1,000,000
At $120 an hour and 238 wpm that is $6,302.52 per million input tokens. At 40 wpm writing, $37,500.00 per million output tokens. Against Claude Fable 5 at $10.00 and $50.00, that is 630 times on input and 750 times on output.
Sources
- 238 wpm, silent reading. Brysbaert, M. (2019). "How many words do we read per minute? A review and meta-analysis of reading rate." Journal of Memory and Language 109:104047. 190 studies, 18,573 participants.
- 40 wpm, typing. Rounded up from 36.2 wpm in Palin, Feit, Kim, Kristensson & Oulasvirta (2019), "How do People Type on Mobile Devices?", MobileHCI 2019, 37,370 volunteers. Two-thumb typists averaged 38. Desktop samples that self-select for typing tests run higher: Dhakal et al. (2018) measured 52 wpm across 168,000 people who chose to take one.
- tokens ≈ words × 4/3. The standard English approximation, roughly 100 tokens per 75 words. Real tokenisers vary by a few per cent.
- $10.00 in and $50.00 out per 1M tokens. Claude Fable 5 list price, checked 19 August 2026, along with the other four in the picker. These move, so they carry a date. Input tokens are priced at the input rate and output tokens at the output rate, item by item, so anything you read and reply to is split rather than billed at whichever rate flatters the argument.
- The machine side is a choice, and the default is not the cheapest. The page starts on Claude Fable 5 at $10.00 and $50.00, the dearest flagship most readers would name unprompted, rather than on an average, because an average is a figure nobody can go and buy. One model in the picker costs more still: OpenAI GPT-5.5 Pro at $30.00 and $180.00, and every multiple here shrinks when you select it. The least flattering version of the argument is one click up the list, which is where an argument worth publishing should keep it. Everything below the default is cheaper on both sides and makes each multiple larger. You are on the default.
- Your hourly rate is not sourced, and cannot be. Both defaults on this page are round starting figures. Every price here is arithmetic performed on a number you supplied, which is why the field sits at the top rather than in a settings panel.
Where this approximates
- Code is not prose. Brysbaert measured silent reading of English non-fiction. A diff is read slower than that, and a diff with unfamiliar code in it is read slower again, so every review figure on this page is an underestimate. That is the same direction the typing default errs in, and deliberately.
- Words are real words. Typing research reports speed in five-character "standard words", which run a little short of average English words. Real words are counted everywhere here for internal consistency, and 40 wpm is used as published. That error is smaller than the error in guessing your own hourly rate.
- Presets state their assumptions. Each one prints how many comments and how many words each, right on the receipt. If yours were different, the paste boxes exist for exactly that.
- Video is not text. An hour of TikTok is billed at an hour of your time, because that is what it costs. The token count is a separate fact about how much content you took in. The gap between the two is the most interesting number on this page.
- Tokens are not comprehension. Reading a page and understanding it are different things, and no rate card can tell them apart. The pass comparison prices passes for the same reason: quality is not on this page because this page cannot measure it.
Prior art
Chris Maury priced human attention in tokens in January 2025, in "Tokens, Attention, and the New Labor Market". He was asking what happens to a career when the unit of work stops being an hour. This asks the smaller question: what it costs you on a Tuesday night, when nobody is buying and you are spending anyway. His method reckons about 6,000 tokens an hour of attention, which makes a human roughly three times more expensive than the reading-speed method used here.
This is not a guilt trip. The point of a rate card is not to spend less. It is to know what you are spending, so that the expensive things are the ones you chose.