Parallax · tech desk · — 18

The Bill Came Due in April

AI keeps getting cheaper per token, and that is exactly why the bill goes up. The price per unit fell, the units per task exploded, and Uber burned its whole 2026 AI budget in four months.

Everyone agrees on the sentence: AI is getting cheaper. Look closer and it is true at the unit level and false at the invoice level, and that gap is the whole story. A scissor has two blades. Watch both.

The blade everyone watches
Price per token keeps falling.
Frontier capability that cost dollars per million tokens in 2023 now costs cents. Claude 3 Haiku launched at $0.25 / $1.25 per million input/output tokens; today's small models sit lower still — GPT-5-mini at $0.05 / $0.40, Grok 4 Fast at $0.20 / $0.50. The rate card has fallen more than an order of magnitude in three years. The corollary feels obvious: an AI budget should be shrinking.
The blade that cuts the budget
Tokens consumed per task climb faster.
The tools stopped being things you talk to and became things that work for you. A coding agent re-reads the whole codebase on every step, calls tools, and runs reasoning loops, so one task can burn well over a million tokens where a chat reply burns a few thousand. Multiply a unit price that halves each year by a unit count that climbs an order of magnitude, and the product, the actual invoice, goes up. Uber budgeted 2026 against last year's cheaper-per-token mental model and exhausted the entire annual figure in four months.

THE SCISSOR

Around November 2025, coding agents crossed from often-work to mostly-work and became daily drivers for highly paid engineers. The shift from a tool you talk to into a worker that runs for you is what changed the token math, because an agent that works autonomously burns vastly more tokens than a chat that answers a question.

Here's the thing the bill doesn't tell you: nobody raised prices. The token bill climbed because, around November 2025, the software quietly changed jobs.

For most of 2025, OpenAI and Anthropic were running what Simon Willison sums up as "Reinforcement Learning from Verifiable Rewards to increase the quality of code written by their models." Basically: teaching the models to write code that passes the tests. The payoff landed in a single fortnight. GPT-5.1 Codex Max shipped on 19 November 2025, Claude Opus 4.5 on 24 November, and paired with their coding harnesses, Willison dates this to the moment coding agents went "from often-work to mostly-work, crossing a quality barrier where you could use them as a daily-driver to get real work done."

A chatbot answers and stops. An agent keeps going. It reads the repository, plans, edits, runs the tests, reads the failures, and tries again. And here's the part that costs money: to hold its place, it replays the entire conversation back to the model on every single step. Willison's own guide puts the mechanics plainly. Coding agents "maintain state by replaying entire conversations with each new prompt," so "as a conversation gets longer, each prompt becomes more expensive since the number of input tokens grows every time." The work that used to be a question is now a loop, and the loop is metered.

And that's the part the cheaper-per-token headlines skip right over. The price of generating a line of code fell to almost nothing. The cost of getting an agent to deliver good code, across a long autonomous run, didn't fall with it. The tools the labs finally found a market for, Willison writes, are exactly the ones that "burn vastly more tokens, but are also quickly becoming daily drivers for the work carried out by extremely well-compensated professionals." The capability showed up, and the consumption showed up welded to it.

THE CROSSOVER

The other blade, plotted. One of Willison's GPT-5 Codex tasks consumed 169,818 input, 17,112 output, and 1,176,320 cached tokens: well over a million for a single task, and the only verified point on this chart, sitting at the top. The vertical axis is logarithmic because that anchor sits three orders of magnitude above the bottom. The lower points trace the shape from a single chat turn upward and are illustrative of task class, not measured.

Tokens per task · chat turn → agentic session · log scale tokens consumed (one task) · log
3,192 2e4 8e4 4e5 2e6 0.68 1.84 3 4.16 5.32 one chat reply* short Q&A thread* single-file edit* multi-step refactor* GPT-5 Codex task (measured) task autonomy → tokens consumed (one task) (log)
Source · Top point: real GPT-5 Codex task (Willison, llm-pricing). Lower points illustrative of task class, not measured.

TOKENS PER TASK

Read top to bottom, the sequence explains itself. The tools got good, so the pricing flipped to metered usage; once usage was metered, the frontier price started rising; and an enterprise budget set against the old mental model ran out a third of the way into the year.

Aug 7 2025
GPT-5 sets the frontier price.
Nov 24 2025
The November inflection.
Apr 2 2026
OpenAI moves Codex to token metering.
Apr 23 2026
GPT-5.5 ships at 2× the price of GPT-5.4.
~Apr 2026
Uber's 2026 AI budget is exhausted.
Jun 2 2026
Uber caps every engineer at $1,500 / month per tool.

HOW THE BILL CAME DUE

The rate card is one number; the realised cost is another. Over thirty days, at standard interface rates, Willison ran up more than two thousand dollars of tokens. He paid two hundred. That gap is what keeps the true cost off the individual's own statement, which is precisely why finance is the last to find out. The dashed line marks what he actually paid.

One heavy user · 30-day token cost at API rates vs. subscription price paidUSD · 30 days
what he actually paid (subscriptions) · 200
Tokens billed at API rates (total) 2180.16
Anthropic Claude Code 1199.79
OpenAI Codex 980.37
Actually paid (Max + Pro) 200
Source · Simon Willison, 27 May 2026 — self-reported 30-day usage

THE REAL PRICE OF A SESSION

Uber is the first named company to run into the scissor in public. The cap is a ceiling, drawn around a line item that, left uncapped, ate a year's budget by April. The figures, set side by side.

UBER · AI CODING SPEND · 2026
$1,500 /mo
Per engineer, per agentic tool
Separate budgets per tool. Applies to Cursor and Claude Code — agents, not chat assistants.
4 months
To exhaust the entire 2026 AI budget
Set against last year's cheaper-per-token, lighter-usage reality.
~11 %
Of a median engineer's total comp
Two tools ≈ $36,000/yr against ~$330,000 comp. Willison's arithmetic on Levels.fyi data, not an Uber disclosure.
$2,180
Tokens one heavy user burned in 30 days
At API rates, for a $200 subscription. The cost the individual never feels.
Sources · Bloomberg (Natalie Lung) via Simon Willison, 2–3 Jun 2026 · Willison, 27 May & 20 Apr 2026

$1,500 / MONTH

That's the short version.

Full issue · 7 sections · 7 sources, all linked

Read the full issue →