Average is dead. It's now free.

The factory that sewed my Lion Cat Hats would have sewn yours. Same pattern, same freight forwarder, same price list, no questions asked.

Nothing about that product was mine. Not the manufacturing, not the materials, not the shipping. The only thing I owned was the reason anyone wanted one, which was a joke about a lion.

I think about that every time someone tells me AI is coming for creative work. It isn’t coming for creative work. It’s doing to competent work what an open factory floor does to a physical product: making it available to anyone, at the price of the raw inputs.

The floor has risen and the middle has collapsed. Competent copy, competent design, competent code. The output that used to cost a trained person a day now costs about as much as the electricity to produce it. If your position in the market was that you do a tidy job, your position has just been repriced to zero.

What it actually costs now

Stanford’s AI Index put a number on the collapse. Hold the standard fixed (GPT-3.5’s score on a standard benchmark) and the price of getting it fell more than 280-fold in the 2 years after ChatGPT arrived. Same output, 1/280th of the cost, no negotiation.

Nothing in business gets 280 times cheaper without something structural breaking. What broke is the price of being adequate.

The 2026 edition shows where it went after that. As of March 2026, 6 labs sit within 25 Elo points of each other at the top of the Arena leaderboard: Anthropic, xAI, Google, OpenAI, Alibaba and DeepSeek. None of them is far enough ahead to charge for it. The competition stopped being about what the models can do and became about what they cost to run.

That’s the second half of the story. Adequate didn’t just get cheap. It got identical everywhere.

Why the middle went first

A language model predicts what comes next given everything that came before. Its output is, by construction, the centre of the distribution. That isn’t a criticism, it’s the design.

So it’s extremely good at producing the average of everything already published on a subject. And the average of everything already published is a fair description of most professional output.

That’s the part people don’t want to sit with. Most of what agencies, mine included, in-house teams and freelancers sold for 20 years was competence. Reliable, correct, unremarkable. It was a real product because it was scarce, and it was scarce because it took years of training to acquire and a day of labour per unit. Both of those costs have gone.

The top didn’t move. Neither did the bottom, because nobody was paying for it anyway. The middle is where the money was, and the middle is what got automated.

This has stopped being a prediction. Employment of software developers aged 22 to 25 is down nearly 20% from its late-2022 peak. In the profession that was meant to be safest, because it was technical. Not the seniors. The entry rung, where the work is competent by definition and remarkable by exception. A third of organisations expect AI to shrink their headcount within the year.

What a model can’t reach

The 2026 report calls it the jagged frontier. A model took gold at the International Mathematical Olympiad. The best of them reads an analog clock correctly about half the time, against 90% for a person. Capability isn’t a line that rises evenly. It’s a coastline, and what’s still yours is in the gaps.

A model can only produce what’s inferable from what’s already public. Which tells you exactly where the remaining value sits. Not in a style. In the things that were never written down:

• What happened to you, specifically, that nobody published.
• What your own data says. Your numbers, your customers, your failure rates.
• A decision made under constraints only you had, with the reasoning attached.
• A position most of your industry would argue with.
• The judgement to know which of 20 acceptable options is the right one.

None of that is talent. It’s all evidence of having been somewhere and paid for the ticket. That’s the whole moat now, and it’s a better moat than competence ever was, because it can’t be trained on until you’ve published it and by then you’re already ahead.

What to do about it

Four things. None of them need a new tool.

Run the inference test on your own output. Take the last 10 things your business produced. Could a capable person with a model and 20 minutes have made this from public knowledge? Where the answer is yes, stop paying a premium for it. Where the answer is no, that’s the part worth protecting.

Stop buying the middle. There are 2 prices worth paying now: the machine price for average work, and the real price for work only a specific person can do. The rate in between is the one being deleted, and it’s the rate most digital work is still quoted at.

Write down what only you know. Your margins by product line. The project that lost money and why. The thing you tried that everyone said would work and didn’t. Most organisations are sitting on years of this and publish none of it. It’s the cheapest advantage available to you and it’s already paid for.

Shift the budget from production to decision. If making things is nearly free, the money belongs with whoever decides what gets made, what gets killed and what gets shipped as-is. That role used to be overhead. It’s now the product.

The uncomfortable version

Some of this will read as a threat if you sell competent work, and it is one. I’ve had the same realisation twice: once when copies of my hats turned up on eBay, and again when I watched a model produce a serviceable landing page in the time it takes to make coffee.

Both times the answer was the same. Own the reason, not the output.

Competence was never the point. It was just expensive enough to pass for a moat.

---

Sources: Stanford HAI, AI Index Report 2026, Technical Performance and Economy chapters. Inference cost figure from the 2025 edition. Stanford Digital Economy Lab, Canaries in the Coal Mine?, 2025. McKinsey, The state of AI, November 2025.