Why I don't subscribe to LLMs

Sep 17, 2026 (Sep 17, 2026) Loading...

In Fences, not Sandboxes, Steve Yegge tells us he burns $120,000 in tokens every month but spends just $4000 on subscriptions... Bargain!!

It's not a bargain: It looks like a confidence trick. It looks like a beat-the-system, too-good-to-be-true, get-it-before-it's-gone confidence trick. It's like picking up a conjoined pair of max-memory M5Ultra MacStudios for the price of the cheapest Neo. I hope Yegge's geared up to hop off the wave before it breaks.

LLM subscription pricing hides the connection between value (what I get) and cost (what it actually costs, not what it costs me). That's not an opportunity, it's a dependency risk – as enterprises recognised earlier this year when their plans switched to API pricing.

So I pay API prices for my LLM usage. The regular chink, chink of money going out helps me to stay aware of my use, helps me compare different ways that I use the tools, helps me – I hope – to manage my dependency. Here's what I've spent.

Aside: Subscriptions...

There's nothing wrong with subscription models: I subscribe to а magazine or a podcast and I pay for each issue; I subscribe to someone's patreon and I support a favoured creator. Some people subscribe to my stuff, too – it's a business model and I'll not knock it.

LLM subscriptions mask the cost, encouraging people to use the LLMs more often, for wider purposes, and to the exclusion of their own skill or the skill of others. Why might they do that? Why might LLM-trainers subsidise individual inventors to invent? Hmm...

Aside: Cost?

For clarity, I'm only really managing monetary cost here. That's not right.

It hides important stuff that we mostly already recognise, in our fevered times: There are so many different costs to using LLMs – financial costs of training and provisioning, job losses and insecurity around sharp technological change, human costs around labelling, societal costs around ownership of content and control of information, environmental costs of compute, inflationary prices as the LLM providers buy up memory and CPU capacity...

Some reckon the end-user per-token cost of inference (i.e. using the tools) can only go down as efficiency grows, use grows, and price-resistance increases. It may – but the costs above won't respond in the same way.