Somewhere along the way, paying $20 a month for ChatGPT Plus and another $20 for Claude Pro became the default. Forty dollars a month, five hundred a year, for tools most people use to write emails and reformat lists. For the vast majority of what you actually do with AI, you are massively overpaying. The frugal move is to default to a cheap model and only reach for the premium one when the task genuinely calls for it.

The cheap models are good now

The Chinese labs have caught up on the things that matter for daily use. Two worth knowing:

  • GLM 5.2 (from Zhipu). Strong general and coding model, available through a subscription that costs a fraction of the American plans. Their coding plan in particular is famously cheap for how much you get.
  • Qwen 3.7 Max (from Alibaba). A capable frontier-class model you can use free in the browser, and the Qwen family also ships open weights, which means you can run smaller versions on your own machine and pay nothing at all.

For writing, summarizing, coding help, data cleanup, translation, and general question answering, these are more than good enough. You will not notice the difference on ninety percent of your tasks, except on the bill.

Default cheap, escalate rarely

The strategy is simple. Make a cheap or free model your default for everything. When you hit a task that truly needs the strongest reasoning available, a genuinely hard problem, a subtle piece of code, something where the best model earns its keep, then pay for one month of the premium tool, do the work, and cancel. You are paying for the frontier only in the rare moments you actually stand at the frontier.

Most people do the opposite. They pay for the most expensive tool every month and use it to rewrite a paragraph. That is the AI version of buying a truck to carry groceries.

Run it yourself for free

The cheapest tier of all is your own hardware. Qwen and other open-weight models run locally, no subscription, no data leaving your machine. I built a whole offline AI chat app around exactly this, a private model on my own device that costs nothing per message and works on a plane. For a lot of daily use, a local model is all you need, and the running cost is zero forever after the download.

The honest tradeoff

I am not going to pretend the gap is zero. For the hardest reasoning and the most demanding coding, the top American models still hold a real edge, and if your living depends on that edge, pay for it, it is worth every dollar. But be honest about how often you are truly at that edge. For most people it is a few times a month, not every day. Match the tool to the task and the savings are automatic.

The move

  • Pick a cheap or free model, GLM or Qwen, and make it your everyday default.
  • Cancel the premium subscription you are paying out of habit.
  • Keep the option to pay for one month of a top model when a hard task genuinely needs it, then cancel again.
  • For private or offline work, run an open-weight model locally and pay nothing.

Five hundred dollars a year for tools you use to reformat lists is not frugal. Default to the cheap model, keep the premium one on standby, and spend the difference on something that matters. More of the same thinking, applied to an entire expensive city, is in Frugal SF.