The Cheapest AI Stack That Actually Ships
You don't need three $20 subscriptions to get real work done. A lean stack beats a fat one.
I audited my own AI spend and found I was paying for overlap. Two chatbots that do the same thing, a video tool I used twice, a research app I opened for one report. Cutting to a lean stack saved me real money and made the tools I kept matter more.
The core is one strong chatbot for writing and reasoning, one specialized model for your heaviest unique task (image, audio, or search), and one automation glue if you move data between apps. Everything beyond that is usually a nicer UI on something you already have.
For cost, the small open models are the quiet win. A $0.04-per-million decision model handles routing and classification that I used to burn a frontier call on. Perplexity's pplx-decider and similar tiny models do the boring 80% for pennies, leaving the expensive model for the 20% that needs it.
The test is simple: if you haven't opened a tool in two weeks, cancel it. A stack that actually ships is the one small enough that you know exactly what each piece is for. Fat stacks feel productive and produce sprawl.