What running your own always-on AI box actually costs.

"Run it yourself" sounds like it could be expensive or unpredictable. It isn't, but you deserve the real numbers rather than a shrug. Here's the real breakdown: three line items, what drives each, and roughly what to expect.

The three things you pay for

When you run your own setup, your bill has exactly three parts — and one company (us) is deliberately not a variable in it.

1. The cloud box

This is the small always-on server that does the work, on a host like Railway, Render, or Koyeb. For a personal box that's mostly idle and springs to life when you use it, this typically lands in the low single digits to about ten dollars a month, depending on the host and how much it's running. It's a flat, predictable hosting cost, the same kind you'd pay for any small always-on service.

2. Your AI usage

The models themselves. You connect your own API keys — Claude, GPT, Gemini, Grok, Perplexity — and you pay each provider directly for what you use, at their price, with no markup from us sitting on top of it. This is the line item that varies with how heavily you work, and it's entirely in your hands and fully in your view.

It's also the line item everyone guesses at, so here's the actual arithmetic. Say you ask 40 questions a day. Call it 1,500 tokens in and 700 out per question — a token is roughly three quarters of a word. That's 60,000 tokens in and 28,000 out per day. Run that against what the providers charge:

Model Per million tokens Your month
Claude Sonnet 5 $2 in / $10 out about $12
Grok 4.3 $1.25 in / $2.50 out about $4.35

Claude's $2/$10 is an introductory rate that runs through 31 August 2026; on 1 September 2026 it goes back to $3/$15, which lifts that same month to about $18. Grok's $1.25/$2.50 is xAI's current short-context rate.

Same forty questions. Nearly three times the bill today, four times once Claude's introductory rate ends. That gap is the thing nobody tells you, and it's the reason an always-on box is affordable at all — the expensive model is only expensive when you actually need it. Ask the cheap one where you don't. Wetlether lets you keep all five connected and pick per question, so you're not paying flagship rates to be told what time it is in Denver.

Check the math. 60,000 in at $2 per million is 12 cents; 28,000 out at $10 is 28 cents; 40 cents a day is about $12 a month, at Claude's introductory rate. The same day on Grok 4.3 is 7.5 cents plus 7 cents, which is $4.35 a month. Prices verified against Anthropic and xAI's own pages on 17 July 2026 and they move — check them yourself before you budget on them. Your questions will not be 1,500 tokens; that's an average to argue with, not a quote.

3. The app

That's Wetlether — the client that ties it together. There are four plans: Desperado is free, Rodeo is $19.99/month, Full Steam is $39.99/month, and Orpheum is $59.99/month. Whatever tier you're on, we never meter you — each monthly price is flat because we run no model and keep no copy of your conversations, so there's nothing on our side to bill by the token.

Put together: a few dollars of cloud, your own AI usage paid straight to the provider, and your chosen Wetlether plan (free to $59.99/month). No metering from us, no surprise middleman markup — the parts that vary are yours to see and control.
The catch. These are ranges, not a quote — your cloud bill depends on the host and your usage, and your AI bill depends on how hard you push the models. What we can promise is the shape: predictable hosting, provider costs at cost, and a plan price that's flat and never metered.

See how it works →

Wetlether — plans from free to $59.99/month, never metered.

Keep reading