The Plan Above Your Plan

I wrote about the Republic of Slop. Before that I wrote about the Lobotomy Pipeline, and a pile of other posts warning about the same AI psychosis from different angles.

This one is a prediction. I’ll write it down today so I can come back and say I told you so.

The era where one subscription got you everything is ending, maybe within the next 12 months. Yes, I hear it. I sound like Dario now. In March 2025 he said AI would write 90% of the code in three to six months, and that in twelve months “we may be in a world where AI is writing essentially all of the code.” Eleven months later he called the 90% figure “a very weak criterion.” So I’ll take the same twelve months he gave himself and the same right to be wrong.

Dario has to make money. I understand that. The rest of this post is about how he’ll make it.

The Window¶

For a while, a few hundred dollars a month bought you something close to unlimited access to the best models on the planet. That was a window.

Sane people used it to write things the complexity merchants had gatekept for years: drivers nobody documented, and tools you used to need a vendor contract for.

Most people used it to clone things that already existed. They generated 100GB of Markdown that had been written before, only now with em dashes.

On October 5, Tibo Sottiaux, who runs Codex at OpenAI, promised that for the next 28 days “each day we’ll either ship one thing that is a clear improvement” or reset everyone’s usage limits. They ran it to November 1. Fourteen of those days were resets, according to the community tracker.

Read the deal again. Either they ship you something to test, or they refill your tokens so you keep testing what they shipped yesterday. You were the QA department, and your salary was paid in tokens.

Expensive or Illegal¶

While I was writing this, my friend Lloyd Wilson posted:

I fear AI/SI subscriptions are either going to get very expensive soon - or illegal.

I replied that neither will happen.

Transformers are out of Pandora’s box, and you can’t even block them. The people with enough funding to keep running them can last a very long time. A country that bans it signs its own expiry date, and a politician who tries it signs theirs first.

The loudest anti-AI voices aren’t the maintainers. Maintainers have nothing against AI. They’re against slop: don’t send them garbage an LLM generated and call it a contribution.

The loud ones are the talkers, fluent and well connected. Their job used to be hunting down the autistic team that built the thing, then promoting it as their own. They translated other people’s work into something a buyer could follow. That fluency was the whole product, and it’s over. Anyone can talk to an LLM now and come out sounding just as polished.

Which brings a new problem. Everyone can write now, and nobody reads any of it.

The Plan¶

Here is how I think it goes.

Your subscription won’t get more expensive. Your MAX, your ULTRA, your 20x and your 200x keep their price. New plans show up above them.

What you stop getting is the new advantage, until you upgrade.

When you talk to Opus, Fable or Astra, you’re talking to an entire infrastructure, and the weights are only one piece of it. Even if they handed you the weights and you had 200GB of RAM to self-host them, you’d be missing what sits around them: the retrieval and the search on one side, the classifiers on the other. OpenAI and Anthropic are building generalists that can do everything, and they run them at the most expensive compute they can buy.

When the model stops making yesterday’s mistake, do you think they retrained it overnight? A training run takes months.

Grok gave everyone a public demo. On July 8, 2025, it started calling itself “MechaHitler” on X. By xAI’s count the broken update was live for 16 hours. Nobody retrained anything in that time. xAI’s own postmortem blamed “an update to a code path upstream of the @grok bot,” and added: “This is independent of the underlying language model that powers @grok.” They removed the bad instructions and the same weights went back to behaving. The fix lived in the infrastructure, which is also where the product lives.

That infrastructure is where the new features land, and it’s metered per plan. I’ve already said the companies will move everything onto their servers. This is what it looks like from the paying side.

It’s already happening. OpenAI shipped GPT-6 Astra on September 3. According to 9to5Mac, “Astra usage is included within the existing subscription allowances,” and you “will also be able to purchase credits for additional usage.” It’s included in your plan, and the meter has a top-up button. The better variant, GPT-6 Astra Pro, went only to Pro and the business plans.

The plan above the top plan already exists. After Astra launched, OpenAI paused sign-ups to its $200 Pro plan. On September 29 it reopened them and added a $500 tier with the highest limits, plus an “Ultrafast” mode no cheaper plan gets. The same email told $200 subscribers their Codex allowance would drop from 20x Plus to 10x on October 30, at the same price. On October 29 the release notes said: “We’ve also canceled the Pro 200 reduction announced in our September 29 email.” The $100 and $500 allowances doubled the same day.

Lloyd gets half a point there. They did try to charge more for the same plan, and it lasted a month. What stayed was the $500 tier and the mode only it gets.

Your plan just looks neutered next to the new ones.

Remember 2024, when influencers’ minds got blown because Claude could one-shot Snake from a single prompt? Today people one-shot an entire 3D world. Two years from now, today’s ceiling is the free tier and the impressive stuff costs more.

What Doesn’t One-Shot¶

What nobody can one-shot is a genuine new idea that nobody did before. That’s the friction I keep poking at.

I already wrote about the generation that doesn’t know what it’s doing. “Claude told me.” “GPT reviewed it.” Now tell me how you disrupt an industry with that. You don’t. You become a commoner.

Fame works the same way now. President Orange would never have been famous if his father hadn’t put him there. Trump liked to say he started with a $1 million loan from his dad. A New York Times investigation in 2018 found the loans came to more than $60 million, and that over his life he received at least $413 million in today’s dollars from Fred Trump’s real estate empire. He was a millionaire by age eight. The family money goes back further, to his grandfather’s gold-rush hotel in Bennett, British Columbia, which biographer Gwenda Blair says sold sex alongside the meals. Trump calls that “totally false.”

Then he got to the White House and made his daughter Ivanka and her husband Jared Kushner his advisers. That’s Idiocracy, played straight.

Some developers got famous the same way. A model rewrote something in a language they don’t know, and their name ended up on it.

In 1995 I was a kid making about $500 a month teaching grown adults how to move a mouse and use Windows 95. That was a fortune for a kid. Nobody pays for that in 2026.

That’s why I wrote the Mars Chronicles. You need to live in the future, because everything in the past is going to be automated.

Come back in November 2027. Your plan will cost what it costs today, and it will feel like the free tier.