Fable 5 Is Back. The AI Coding Market Already Moved On.

On June 12, the US government told Anthropic to pull Claude Fable 5. Three days after it launched. The order said no foreign nationals could touch it, including Anthropic’s own foreign employees, and since nobody could verify nationality on every request, the model went dark for everyone. API, Claude.ai, Bedrock, all of it. Gone.

On June 30 the controls came off. On July 1 it was back in your editor.

Three weeks offline. National security. A frontier model too dangerous to export.

Here’s the part that got buried. Amazon researchers found a way to get Fable 5 to flag software vulnerabilities, and in one case show how to exploit one. So Anthropic tested that against a stack of weaker models. Opus 4.8 found the same bugs. So did GPT-5.5. So did Kimi K2.7, a Chinese model you can download for free. Every model they tried could reproduce the exploit demo. Haiku 4.5 could do it.

We spent three weeks treating one company’s model like enriched uranium. The capability was already sitting on Hugging Face.

That is the state of the AI coding market in a single story. Everyone is guarding a frontier that stopped being scarce.

The frontier is cheap now

Look at where the tokens actually go.

By May, Chinese open-weight models were around 61% of everything routed through OpenRouter, the largest neutral model router. Four of the five most-used models were Chinese. Meta’s Llama, the open-weight leader eighteen months ago, fell off the list entirely. And more than half of that traffic is code. People aren’t running these models to write poems. They’re shipping software with them.

In one seventeen-day stretch this spring, four Chinese labs dropped open coding models that trade blows with the Western frontier. GLM-5.1. MiniMax M2.7. Kimi K2.6. DeepSeek V4. None of them costs more than a third of the closed US models. DeepSeek runs around thirty cents per million tokens. Claude runs ten dollars. Same task. Thirty-five times the price.

A part-time developer in San Diego told Rest of World that an hour of coding cost him ten dollars on Claude and under fifty cents on DeepSeek. The founder of Lindy switched off Anthropic and said it saved his company millions. His line was better than anything I could write: you don’t need God to write your email.

That is the real competitive pressure. Not a leaked frontier weight. A cheap, good-enough model that does most of the work for a tenth of the price.

The export controls were fighting the last war.

The money says the tool is the prize

Meanwhile the valuations went vertical.

Cursor crossed two billion dollars in annual revenue by February and is raising at fifty to sixty billion. SpaceX took an option to buy the whole company outright for sixty billion, or walk. It’s the fastest-scaling B2B software company on record. Zero to two billion in under three years. Slack took seven years to reach its first billion. Cursor did it in eighteen months.

Cognition, the Devin people, raised at twenty-six billion on under five hundred million in revenue. That’s a richer multiple than Cursor, which earns four times more. The market is betting the autonomous agent has a bigger ceiling than the assisted editor.

Even the hype turned over. In February, Andrej Karpathy called vibe coding passé. The same guy who coined the phrase a year earlier. Twelve months from meme to obsolete. That’s the clock speed we’re running at.

Read those numbers and you’d think the tool is the prize.

It isn’t. The tool is becoming a commodity feature. Microsoft bundles Copilot into VS Code at basically no marginal cost. Anthropic ships Claude Code on its own API. Google and Amazon can hand out coding agents as loss leaders to sell more cloud. When every environment ships a competent agent, best editor is a lead you hold for one release cycle. Cursor is winning right now because it has the best product. Best product is the most fragile moat in software.

Seventy percent of engineers already run two to four of these tools at once. Nobody is loyal. They route the problem to whatever wins the problem.

What actually changed for the people doing the work

Strip out the funding rounds and the export drama. What moved on the ground?

Adoption is over as a question. Ninety-five percent of engineers in the Pragmatic Engineer’s 2026 survey use AI tools every week. Three quarters use them for at least half their work. Nobody is asking whether anymore. They’re asking which.

But throughput did not jump the way the pitch decks promised. METR ran the cleanest test we have and found experienced developers were 19% slower on real tasks with AI, while they believed they were 20% faster. DORA found the same shape at the system level: individuals feel faster, delivery gets worse. Everyone ships more code. Not everyone ships more working software. Those are different things, and the gap between them is where this year actually lived.

The jobs picture split down the middle. Entry-level postings are down somewhere between 30 and 40% from the 2022 peak. Senior and staff roles are growing. A senior engineer with Claude Code does the work of a small team. That same tool in the hands of someone who has never designed a system produces confident, broken code that falls over in production. So the market pays up for the people who can tell the difference, and stops hiring the people who were on their way to learning it.

Which is the problem everyone saw coming and walked into anyway.

Gartner says half the companies that cut jobs for AI reasons will rehire for the same roles. Forrester says 55% of employers already regret those layoffs. A third of them spent more restaffing than they saved. Klarna went all in on replacing people, then started hiring them back. Salesforce hired zero engineers in fiscal 2026. And the same firms cutting seniors are quietly reposting junior reqs six months later, because the codebase is a mess and nobody junior was around to learn it.

We ran this exact experiment after 2008. Freeze junior hiring, and three years later you can’t find anyone with three years of experience. AI didn’t repeal that. It handed everyone a better excuse.

And the Bureau of Labor Statistics still projects software developer employment growing 15% through 2034. The death of the job got announced again. It keeps not happening.

The part that didn’t move

Here is what a month of very loud headlines did not touch.

The model got banned and came back. The frontier got cheap. A code editor became worth sixty billion dollars. And the actual job, the thing you get paid for, is exactly what it was.

Someone still has to know what to build. Someone has to read the code the agent wrote and catch the subtle wrong thing before it reaches a customer. Someone has to own the system at 3am when it goes down, understand why it went down, and decide what to do next. AI made writing code cheap. It made none of that cheap. If anything it made it worth more, because there is now ten times as much confident, plausible, untested code flowing toward production, and someone has to stand at the gate.

The tools changed again this year. They’ll change again next year.

The job didn’t move.

Final Words

I could be wrong about the timeline. Maybe the agents get good enough next year that the gate I’m describing shrinks to almost nothing. I don’t think so. But I’ve been doing this long enough to have been wrong before, and I’d rather hear the counter-argument than pretend it doesn’t exist. If you think a scarce frontier still matters more than I’m giving it credit for, tell me why.

You can find me on LinkedIn at linkedin.com/in/ivanturkovic, on X at x.com/ithora, and on Threads at threads.com/@ithora. That is where most of these arguments actually get tested.

If you want to talk any of this through in the context of your own team, my contact page is at ivanturkovic.com/contact.

And if you are a founder or CTO trying to work out where AI genuinely helps your engineering org and where it quietly hurts it, you can book a call at cal.eu/ivan-turkovic/30min.

So here is my question. If a frontier model can vanish for three weeks and your team barely notices because the cheap open one covered the work, what were you paying the premium for?