01 · THE ARITHMETIC
Cheaper tokens, bigger bills
What most people think
AI gets cheaper every year, so a company's AI bill in 2026 should be smaller than its bill in 2025.
What the data shows
Uber set aside a whole year of AI money for 2026. It ran through all of it by April, then capped every engineer's spending.
Cheaper per unit, many more units per job. That hisaab only goes one way.
Uber exhausted its entire 2026 AI budget in four months and capped each engineer.
Source · Bloomberg (Natalie Lung) via Simon Willison's Weblog · 3 June 2026
02 · THREE WORDS
Only one of these is your bill
Look at your electricity bill. There is a rate per unit, and there are units used.
token. A piece of a word. It is the unit AI companies price their work by.
per-token price. What one of those pieces costs. In 2023 the cheap models started at $0.25 per million, and capability that cost dollars then now costs cents.
cost per task. The price of one piece times how many pieces the job used. On your electricity bill, this is the bottom line.
Source · Simon Willison's Weblog · 4 June 2026
03 · TOKENS PER TASK
One job, over a million tokens
So how many of those pieces does one job use? A chat reply is a few thousand of them. One measured coding job used over a million. On this chart, each step up the side is ten times the one below it.
One measured coding task used 1,363,250 tokens. A chat reply uses a few thousand.
Source · Simon Willison's Weblog, llm-pricing. Lower point illustrative, not measured · 4 June 2026
04 · WHAT CHANGED
The software changed jobs
In November 2025 the software stopped answering questions and started doing whole jobs. A job is a loop.
The tools got good. In November 2025, two coding models shipped a week apart. Coding agents went from often working to mostly working.
A job is a loop. The agent reads your code, tries a change, runs the tests, reads the failure, and tries again.
It re-sends everything. The developer Simon Willison, who tracks AI pricing, writes that agents "maintain state by replaying entire conversations with each new prompt".
Source · Simon Willison's Weblog · 19 and 27 May 2026
05 · INSIDE ONE TASK
Most of it was re-reading
So the loop re-sends everything it already has, on every step. Here is one real coding task, split three ways. Cached means the pieces the agent had already sent once.
One measured task used 1,363,250 tokens. Of those, 1,176,320 were cached. Fresh input was a far smaller slice. Only 17,112 were new code.
Source · Simon Willison's Weblog, llm-pricing · 4 June 2026
06 · HOW IT HAPPENED
Billing changed, prices rose, budget gone
Almost all of that task was context the agent re-sent. Now watch what the billing did next. Frontier means the newest and most capable model on sale.
- Nov 24 2025
Coding agents start mostly working. Nothing got more expensive here. The tools just started working.
- Apr 2 2026
OpenAI starts billing Codex by the token. Priced on token usage now, not per message.
- Apr 14 2026
Anthropic starts billing the seat by usage. A seat is one person's licence. About $20 a month, plus what you use.
- Apr 23 2026
New frontier model, twice the price. At the frontier, the per-token price is now rising.
- ~Apr 2026
Uber's 2026 AI budget is exhausted. Four months in, the whole year's money is gone.
- Jun 2 2026
Uber caps engineers at $1,500 per tool.
Source · Simon Willison's Weblog · 19 May to 3 June 2026
07 · THE CAP
One engineer, one tool, one month
So the cap arrived on 2 June, one limit for each engineer on each tool. It covers only the tools that write and run code on their own.
The cap covers agent tools like Cursor and Claude Code, not chat assistants. Over a year on two tools, the cap adds up to tens of thousands of dollars. The pay beside it is an outside estimate, not Uber's own number.
Source · Bloomberg (Natalie Lung) via Simon Willison's Weblog. Levels.fyi for the pay estimate. Converted at about ₹95 to the dollar, September 2026 · 3 June 2026
08 · THE BILL ELSEWHERE
Where else the money moved
So one engineer with two tools can spend $36,000 a year. Here are four more readings, taken off other meters.
Tokens one heavy user burned in 30 days: $2,180. Valued at the advertised price. The subscriptions behind it cost $200.
Newest top model's price against the last one: 2×. Shipped 23 April 2026. At the top end, prices are rising.
Cost rise with the price per piece unchanged: ~1.4×. The April update counts the same text as more pieces.
Of a median engineer's estimated total pay: ~11 %. Two tools a year, against an outside estimate, not Uber's figure.
The token bill, four readings
Source · Simon Willison's Weblog. Levels.fyi for the pay estimate · 20 April to 3 June 2026
09 · WHAT ACTUALLY MOVED
The work changed shape
Those four numbers point the same way. One price rose in April, at the top of the market. Where the price per piece held still, the bill rose, because the work changed shape, and a business pays for work. Think of it like your electricity bill again. The rate per unit fell, which means very little when one job went from a few thousand units to over a million. Writing new code is close to free now. Getting good code is still expensive, and that is what the meter counts.
Source · Simon Willison's Weblog