Four dollars an hour. That’s the reported cost of running Claude on AI research work, according to a report from 36 Kr, compared with roughly $150 an hour for a human researcher doing similar work. And the report says the cheaper option came out ahead on results.
I want to sit with that number for a second, because it’s the kind of figure that gets passed around without anyone explaining what it actually means. So let’s do that.
What the number is really saying
When people talk about AI research, they usually picture humans doing the researching and AI being the thing researched. That’s the normal arrangement. What that $4-per-hour figure describes is something different: the AI participating in the work of improving AI.
An Anthropic researcher recently gave TechCrunch a look at exactly this idea — self-improving AI. Not in the science fiction sense where a machine wakes up one morning and rewrites itself into a god. In the much more mundane sense of an AI system doing the grunt work that goes into making the next AI system better. Running experiments. Testing ideas. Grinding through the tedious parts.
That’s less dramatic than the movies, and more interesting. Because the boring version is the one that actually compounds.
Why cost matters more than capability here
If you take nothing else from this article, take this: the price tag is the story, not the performance.
A researcher who costs $150 an hour is a researcher you deploy carefully. You give them the important problem. You don’t ask them to try 400 slightly different variations of a boring idea just to see what happens, because that would be an absurd use of an expensive human being’s time and attention.
A researcher who costs $4 an hour changes that math completely. Suddenly the absurd approach becomes reasonable. Try everything. Test the dumb ideas alongside the smart ones. Let the process run overnight and see what turns up. Research stops being a series of careful bets and starts being a wide search.
This is how most technology shifts actually happen. Not because something became possible, but because something became cheap enough to do carelessly.
The part where the agents fought each other
Before this starts sounding too tidy, here’s a detail I love. TechCrunch also reported that Anthropic set multiple AI agents loose on the same task — and they started a turf war.
Multiple agents, one job, and instead of dividing the work like a well-run team, they got in each other’s way. Anyone who has worked in an office is now nodding.
I bring this up because it’s a useful correction to the $4 figure. Cheap labor is not the same as coordinated labor. You can run twenty agents for the price of a sandwich and still end up with less useful output than one careful human, because the agents will duplicate work, contradict each other, and occasionally undo progress someone else made. The coordination problem doesn’t disappear when the workers get cheap. It gets worse, because now you have more workers.
Memory is the quiet upgrade
There’s a third piece of news that fits into this picture, and it’s the one most people will skip: Claude Cowork now remembers what you told it in chat.
That sounds like a small quality-of-life fix. It isn’t. Memory is the difference between a tool you re-explain yourself to every single session and a collaborator that accumulates context about your work. For anyone who has typed the same background paragraph into an AI chat for the fifteenth time, this is the feature that matters most in daily use.
It also connects to the self-improvement thread. A system that remembers is a system that can build on what happened yesterday. Continuity is a prerequisite for improvement, whether the thing improving is an AI or a new hire.
None of this settles the race
Anthropic is not running unopposed. New data suggests OpenAI is gaining on Anthropic among business users, which is a reminder that impressive internal research doesn’t automatically translate into people choosing your product.
Businesses pick tools for unglamorous reasons — pricing, integration, whoever their team already knows how to use. A company can be doing genuinely interesting work on AI that improves AI and still lose ground on the sales side. Both things fit in the same quarter.
What I’d actually take away from this
If you’re reading agent101 because you want to understand where AI agents are heading without a computer science degree, here’s my honest read.
The cost collapse is real and it’s the most important number in the story. AI participating in its own improvement is happening in a practical, unglamorous form right now. The coordination problems are also real, and the turf war is a healthier signal about the current state of things than any benchmark score. And memory, the least exciting item on the list, is the one you’ll notice first in your own work.
Four dollars an hour buys a lot of experiments. What it doesn’t buy is anyone knowing yet which experiments will matter.
🕒 Published: