The Thirty-Pence Finish Line

There is a particular smell in the room whenever someone tries to sell you the ethical alternative to a Silicon Valley product. It is the smell of cold steamed broccoli.

You know the pitch. A dinner party, a lukewarm lager, and someone with a canvas tote bag explaining that their phone runs a de-Googled Android, backed up to a Nextcloud server on a Raspberry Pi in the airing cupboard. They have the serene, suffocating piety of the hybrid drivers in South Park, the ones who bend down to enjoy the smell of their own farts.

“It doesn’t send your location to anyone,” they say.

“Right,” you say. “Can I pay for my parking on it?”

“Well, no. The parking app won’t log in. There’s a workaround: you register on the website first, then deny it your location the first time it opens, and if that doesn’t work you flash a build with microG and spoof the integrity check, and…”

And the soul leaves your body.

I can say all this with love, because I spent an evening this week doing exactly what the tote bag did: a de-Googled phone maker in one tab, an encrypted-email company in the other, comparing them like a man choosing a conscience. I know the smell from the inside. And here is what the evening taught me. The privacy movement sells negative space. It tells you what the software doesn’t do. It doesn’t track you. It doesn’t sell your profile. It doesn’t phone home. All true, all admirable, and none of it is a reason to buy anything, because nobody wakes up on a Tuesday wanting digital sovereignty. They wake up wanting to buy a train ticket, send an invoice and look at pictures of their niece without the screen freezing. A phone that honours your rights and won’t open your banking app until you switch off half of its own security settings is not a moral victory. It is an expensive paperweight with a good heart.

Compare the one ethical gadget that sells: the Fairphone. Nobody needs a philosophy degree to understand a Fairphone. You drop it on the stairs at Clapham Junction, the screen goes, you order a new one from the maker for about £85, and you fit it yourself in twenty minutes with the little Torx screwdriver they sell for a tenner. iFixit, which takes phones apart for a living, has scored every Fairphone since the second one ten out of ten for repairability. The company promises software until 2033, shipped 42% more phones last year in a shrinking market, and is closing in on its millionth. The ethics come free with the solved headache.

Keep that pattern in mind, because the thing that wins is always the thing that finishes the boring bit. Now let me tell you about the most profitable sentence in the history of software.

And There You Have It

I don’t write code. I can’t read much of it either. I have spent the last two years building a software company anyway, and the thing I am still angriest about is a tutorial.

Everyone who has ever tried to make something without a developer has sat through it. It is called Build a Complete Web App from Scratch and it is forty-five minutes long. You install things you don’t understand. You type twenty lines of JavaScript you don’t understand. You sweat over a missing semicolon. And then the instructor types the sacred closing lines:

.then(result => {
  console.log(result);
});

They flip to a hidden drawer in the browser, point at a grey blob of text that says { status: 200, success: true }, and announce: “And there you have it! Our app is complete!”

No, it isn’t. There is no screen. There is nowhere for a customer to type their name. Nobody has said what happens if they press the button twice. That is not an app. That is a till receipt.

It took me an embarrassingly long time to learn what console.log(result) means. It is a sign on the door: the tutorial ends here. The rest of the app, the part a customer will touch, is your problem now. To the developer, a message printed to a hidden terminal is a finish line, because the plumbing didn’t explode. To anyone trying to earn a living, it is scrap paper. Stopping there is a builder tipping a pile of wet bricks onto your front lawn, brushing off his hands, and saying: “Right then, that’s your extension sorted. Don’t forget the five stars on Checkatrade.”

The internet has a name for this. Draw two circles, then draw the rest of the owl. Only here it has three steps. Draw two circles. console.log(result). Now draw the rest of a working business.

The people who sign up to learn this way mostly never get past the circles. When Harvard and MIT counted everyone who had registered for one of their online courses, the share who finished had fallen to about three in a hundred by 2018, and it had been falling for years. The tutorials did not fail those people. The tutorials were never going to the end.

The Expensive Bit Was the Boring Bit

The lazy version of this argument says the developer class ran a thirty-year racket, convincing every baker and dry cleaner that writing software was alchemy. I would love that to be true, because it is a better story. The numbers say something smaller and more interesting: most bakers never bought the alchemy at all.

Two in ten UK businesses still have no website. Among sole traders it is one in four. Of the businesses that do have one, only a quarter hand it entirely to an outside developer or platform; 44% run it themselves. In America, where someone bothered to ask how the sites were built, one small-business website in eight was written in actual code. The rest came from Wix, Squarespace, WordPress and their cousins. Nobody fleeced the baker. The price kept him out of the room before the conversation started, and the baker, sensibly, made do with a template and a Facebook page.

Because here is the price. A contract software developer in Britain bills a median of £525 a day; an agency, more. For that money you expect the clever bit. And the clever bit, it turns out, was never where the money went.

Ask the people who build software what they spend their days on and the answer is not architecture. It is the last mile: the button that is clipped on an iPhone, the form that loses what you typed, the thing that breaks when a customer presses it twice. When Anthropic went through what people ask its models to do, the single most common task on the whole platform was “modifying software to correct errors”. Fixing bugs: one in ten of everything coming through its API. Coding as a whole is roughly a third of conversations on Claude.ai and just under half of API traffic, and most of it is the owl.

That boring mile was the moat. It was never guarded by brilliance. It was guarded by the price of a human being’s patience: someone willing to look at a screen, notice the button is clipped, fix the CSS, rebuild, look again, and bill you £525 a day for the privilege. Patience was the expensive ingredient. Then patience got a price list.

What a Look Costs

In March 2023 the best model you could rent, GPT-4, charged $60 for every million tokens it wrote (a token is about three-quarters of a word). One pass of the boring loop, feeding it the page, a screenshot and the stylesheet and getting a fix back, is about 30,000 tokens in and 2,000 out: call it a dollar a look. Fifteen looks an hour, fifteen dollars an hour, and a developer could fold their arms and sneer at the idiot without a degree whose agent had got itself stuck in a loop. What a waste of money.

Then the price fell off a cliff. By August 2024 the same OpenAI flagship charged $10. Google’s small model went to 30 cents that month and to 15 cents in October, the cheapest a big lab has ever sold a million tokens. In November 2024 the venture firm a16z measured the trend and named it LLMflation: for a model of the same ability, the price was falling ten times a year, a thousandfold in three. Epoch AI checked again in March 2025 and found prices falling anywhere between nine and nine hundred times a year depending on the task, with a median of fifty. In January 2025 DeepSeek put out a reasoning model at a twenty-seventh of OpenAI’s price for the same tier, and Nvidia lost $593bn of market value in a day, the largest one-day loss any company on Wall Street had ever had.

Here is the arithmetic that matters, at this week’s prices and this week’s exchange rate. One look at the screen, 30,000 tokens in and 2,000 out, costs:

• about 0.3p on the cheapest capable models (Gemini 2.5 Flash-Lite, Claude Haiku 5.5, OpenAI’s GPT-6 Luna);
• about 6p on a mid-range model (Claude Sonnet 5.5, at $2 in and $10 out);
• about 30p on the dearest models on sale, Claude Fable 5.1 and GPT-6 Astra, at $50 a million out.

Thirty pence is the finish line at its very dearest: the best brain money can rent, looking at your screen, noticing the button is clipped on an iPhone, swearing quietly in binary, rewriting the CSS, rebuilding the preview and looking again. A third of a penny if you aren’t showing off. Against £525 a day for a human who will put it on next sprint’s backlog.

And here is the part the “AI gets cheaper every year” crowd will not enjoy, because I didn’t. The floor stopped falling two years ago. Google withdrew its 15-cent model in June, and the cheapest thing any of the big three sells today is 40 to 50 cents a million, about where it was in the summer of 2025. At the top, prices have gone back up: OpenAI’s flagship was $10 a million in August 2025, $30 in April and $50 by September, and Google has already published the rises it has planned for December and January. LLMflation was a 2023-to-2025 phenomenon. The chart has an elbow in it.

It doesn’t matter, and that is the whole point. The last mile never needed the best brain. Noticing a clipped button is not PhD work. The boring bit, the moat, got cheap enough to brute-force in 2024 and has stayed there, while the clever bit, the bit the priesthood charged for, is quietly getting dearer again. The economics of software turned upside down, which is a different thing from getting cheaper. Trial and error at a third of a penny a go just beat thirty years of architecture.

The Developers Have a Point

They do, and it deserves its say.

Stack Overflow’s 2025 survey of 49,000 developers found 84% using or planning to use AI tools, and more of them distrusting the accuracy of what came back (46%) than trusting it (33%). The frustration two-thirds of them named first was “AI solutions that are almost right, but not quite”. In early 2025 the research group METR timed sixteen experienced open-source developers on real tasks in their own codebases, with and without AI. With it they took 19% longer, and afterwards still believed it had made them 20% faster. (METR’s follow-up this February, with newer tools and 57 developers, found them a little faster and called its own evidence weak.) Brute force leaves marks, too: an agent that clicks until it works can click itself into a security hole it will never notice, and no amount of looking at the screen tells you whether the payment code is safe.

All true. But look at what the complaint is. “Almost right, but not quite” is a precise description of every tutorial ever recorded, and of most first drafts by most humans. It is the state of all software at four o’clock on a Thursday. The question was never whether the first attempt is right. It is who gets to make the second attempt, and the third, and what each one costs. For thirty years the answer was a professional, at a day rate, after you had waited your turn. Now it is anyone, at a third of a penny, immediately, and the machine checks the thing the customer actually sees, which is more than the tutorial ever did.

The developers are objecting to the quality of the owl. Fine. The point is who is allowed to draw one.

Press Enter

The developers, to their credit, have noticed which way the wind is blowing and joined the queue. In February 2025 Andrej Karpathy, who used to run AI at Tesla, described “a new kind of coding I call vibe coding, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists”. By November, Collins had made it word of the year. By this month, two-thirds of the developers in Stack Overflow’s new survey were using coding agents themselves. Lovable, the Swedish company that lets people who can’t code type a sentence and get a working website, went from its first million in revenue to a $100m annual run-rate in eight months, and passed $600m last month, with a million new projects a week being built on it.

So here is what I would tell the version of me who sat through the tutorial. When the video stops at the grey blob, the tutorial is over and the brief has just started. Type the thing the instructor never did: show me the screen. Then keep going. Where does the customer put their name? What happens if they press it twice? Show me on a phone. Each look costs thirty pence at the very most, and the machine does not sigh.

My own company, for what it’s worth, is nothing cleverer than that. There is not a line of technical innovation in it: a hosting account anyone could open, a folder of plain-text files, and a very stubborn cat in sunglasses who keeps clicking the preview until the button works. Two years of being told from the sidelines that this wasn’t possible, by people who later forwarded me a demo of someone doing it. Serenity now.

Which brings me back to the broccoli. The tote-bag phone and the tutorial fail in exactly the same place: the last ten feet, where a human being needs to pay for parking or press a button. The Fairphone sells because someone finished the boring bit, right down to the screwdriver. The software that wins will be the software whose makers were too stubborn, or too cheap, to stop at the receipt.

The priesthood had a good run. The cat is out of the bag, he is wearing sunglasses, and he does not care about your syntax errors. And there you have it.

This post started as an evening comparing de-Googled phones and a long argument with Gemini about tutorials. Data referenced from the DSIT UK Business Data Survey 2026; Clutch, “The State of Small Business Websites” (2025); ITJobsWatch contract rates (October 2026); Reich & Ruipérez-Valiente, “The MOOC pivot”, Science (2019); the Anthropic Economic Index (January 2026); the published price lists and dated announcements of OpenAI, Google, Anthropic and DeepSeek (2023–2026), converted at the ECB’s rate of 14 October 2026; a16z, “Welcome to LLMflation” (November 2024); Epoch AI, “LLM inference prices have fallen rapidly but unequally across tasks” (March 2025); Reuters on Nvidia (27 January 2025); the Stack Overflow Developer Surveys of 2025 and 2026; METR (July 2025 and February 2026); Andrej Karpathy on X (February 2025); Collins Dictionary (November 2025); TechCrunch on Lovable (2025–2026); and Fairphone’s and iFixit’s own pages. The opinions, and the two years of swearing at tutorials, are my own.