Two results pull the other way. On Ramp Labs' 137-task accounting benchmark, no model fully solves more than 21.2% of tasks in three attempts. And in Ramp engineer Sandra Yang's BugBasher experiment, an agent given a card, a phone and an email spent about $3,000 over four weeks to earn $75.6,7 Both readings fit this post's conclusion: the hour is cheap on well-specified work, and everything outside the specification is still where the cost lives.
For three years the argument about AI and work has been conducted almost entirely in adjectives. This week a16z put numbers on it.
Alongside the price, the capability. On the standard benchmark for an agent operating a real desktop, agents now complete 85% of tasks against a human rate of 72%. A year ago the best model managed 42%. a16z's summary: in the last eighteen months, computer-use agents crossed from demo to deployable.
Before going further, handle the numbers honestly. The $6 to $8 agent figure is a founder estimate, cross-checked against frontier token pricing and published inference economics, with a stated range of $3 to $15. The $10 offshore figure comes from real 2026 outsourcing rates. So this is a carefully built estimate meeting a hard market price. It is not an audit, and anyone quoting it as one is overreaching.
Even discounted heavily, the direction is not seriously in dispute.
So why has nothing happened
Here is the part that should stop you.
Since ChatGPT launched, employment in the Philippines' IT and business process outsourcing industry, which is 8% of that country's GDP and by this logic the single most exposed labour market on earth, has grown 20% to 1.9 million workers.
The most automatable work, in the most cost-sensitive market, now priced above the alternative. And the headcount went up.
That is the same shape as the firm-level hiring data covered in the jobs post: the number that should be falling is not falling. So the crossover is real and the substitution is not, at least not yet, and the space between those two facts is where the actual work lives.
Why an agent hour is not an hour
The comparison quietly assumes the two units are interchangeable. They are not.
A person absorbs ambiguity. You can hand someone a half-specified job and they will find the missing context, ask the right colleague, notice the thing that was not in the brief, and stop when it feels wrong. None of that is in the $35.
An agent needs the work decomposed, the tools connected, the credentials scoped, the failure path defined, and someone who notices when the output is confidently wrong. None of that is in the $6.
So the honest comparison is not $6 against $35. It is $6 plus the cost of making the work legible to a machine, against $35 for someone who makes it legible themselves.
On a task you run twice, the person wins easily. On a task you run ten thousand times, it is not close. Which is exactly why the crossover shows up first in high-volume, well-specified, verifiable work and nowhere else, and why it has not shown up in aggregate employment. Most work is not that.
What to do with the number
Stop pricing the hour. The unit that matters is cost per completed task, because it is the only one comparable across a person and an agent. An hour of anything is an input, and you do not buy inputs.
Find the work that is already legible. Clear input, checkable output, runs constantly. The economics there already favour an agent and have done for months. You do not need a strategy for this, you need a list.
Budget the decomposition, not the inference. The expensive part of a $6 hour is everything you do before it: defining the loop, wiring the tools, deciding what happens when it is wrong. That cost is one-time and it is where projects actually die.
The crossover happened quietly and the labour market has not noticed. That is not because the numbers are wrong. It is because a cheap hour is worth nothing until someone has specified what to do with it.
The hour got cheap. The specification did not.
Keep reading.
Teardowns, benchmarks and playbooks on building AI-native products.
Browse the Playbook →