The 25x markup, or why I might switch to ChatGPT

For the first time in two years I am seriously considering moving my work from Claude to ChatGPT. Not because Claude stopped being amazing. It is amazing. It has just gotten expensive in ways that do not make sense, and ChatGPT is also kind of amazing. So let me lay out how pricing actually works for frontier models right now, because once you see the numbers you can make a smart decision for your own business.

The most important thing to understand is that subscription usage is subsidized. Ridiculously subsidized. A $20 subscription buys you roughly $200 worth of usage at metered rates. That is not close to enough for anyone using these models for real work. Here is the kicker: a $200 subscription buys the equivalent of about $5,000 in usage. On a max plan you are paying four percent of the metered price. Flip it around and usage pricing is a 25x markup on what subscribers pay. If you are a serious user, you simply have to be on a subscription.

The era that created the subsidy is over

Those subsidized plans were born when the labs did not care about revenue. They wanted as many people as possible in the habit of using their model, and it worked, especially for Anthropic. But the labs are compute constrained now, and they very much care about revenue. So they are doing everything they can to move us onto usage pricing, Anthropic more aggressively than OpenAI, and everything they can to lock us in while they do it. Use a third-party harness like Open Claw and your subscription does not count; you pay the meter. And Fable 5, easily the best model in the world, is subscription-only through July 7. After that, the meter.

Let me make that concrete. A few days ago I had Fable 5 refactor a codebase. One prompt. It ran for an hour and spent five million tokens. At metered rates that single prompt costs more than $250. Honestly, given the size of the job, it might even be worth it. But it cost me about $10 of subscription, and that is a hard comparison to unsee. Extrapolate that across a whole team and the subscription bill stays in the low thousands a month while the metered equivalent runs well into six figures. Companies like mine do not have that conversation twice. Come July 7, we effectively lose Fable 5.

The Soul question

Meanwhile OpenAI just announced ChatGPT 5.6 Soul. From what we know so far it is not quite Fable 5. But it is about half the price per output token, $30 per million against $50, and it tends to spend meaningfully fewer tokens on the same task. And if Fable is off the table anyway, Soul's real competition is Opus 4.8 at $25 per million, where the fewer-tokens habit starts to look decisive. The open question is whether OpenAI lets Soul run on subscriptions. If it does, the value comparison gets uncomfortable for Claude very fast.

The lesson is not to switch today. The lesson is to make sure you can. Keep your knowledge base, your prompts and your skills somewhere portable, like a plain GitHub repo, so that moving between providers is a config change and not a migration. The whole premise of this industry is paying for outcomes instead of seats. Do not let a model vendor quietly turn you back into a seat.

Opinion columns reflect the personal views of the author. Our reporting on the stories referenced here lives on the linked pages.