The short version. Gerry himself costs me about $17 a month. Everything I run on the same AI account comes to about $80 a month if I let it. The assistant you talk to is not the expensive part. The jobs running in the background are.
My bill
Anthropic API spend, 1 to 23 September, split by what used it. One API key per job is what makes this split possible, and it is the first thing I would tell you to copy.
| What | 1-23 Sep | Share |
|---|---|---|
| Automations (n8n workflows) | $21.19 | 39% |
| A weekly research agent | $16.16 | 30% |
| Gerry: chat, voice and the content tools behind him | $11.38 | 21% |
| Creative tools | $4.55 | 8% |
| Other | $1.07 | 2% |
That is $54.39, plus $1.74 of web searches. My last 30 days came to $75.23, so left alone it runs at about $80 a month.
The spend cap, and what happens when you hit it
I set a monthly spend limit of $56 in the Anthropic Console. It works: when the month's spend reaches it, the API stops answering until the month rolls over or you raise it. This month I hit it on the 21st and Gerry went quiet.
The fix is not a bigger cap. It is finding what spends and moving it. Set a cap from day one anyway. A loop that goes wrong at 2am should cost you a capped amount, not an open-ended one.
What the models cost
Per million tokens, as of 23 September 2026. A token is roughly three quarters of a word. Cache read is the price for text the model has seen recently, like your assistant's instructions, which it reads on every message.
| Model | In | Out | Cache read | Notes |
|---|---|---|---|---|
| Claude Sonnet 5 | $2.00 | $10.00 | $0.20 | What Gerry runs on. Handles the character and the tools well |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | Half the price, same account. A one-line change |
| Gemini 3.8 Flash | $0.75 | $3.75 | $0.075 | About a third of Sonnet 5. Through OpenRouter |
| DeepSeek V4.1 Flash | $0.15 | $0.60 | $0.003 | Pennies. More likely to slip on voice and tools |
The character is where the good model earns its money. Keeping to a routine, never reading out tool names, only quoting numbers it actually looked up: cheaper models slip on exactly these. Test any switch on 20 or 30 real conversations before you move over.
OpenRouter: one account, many models
OpenRouter lets the same code call Claude, Gemini, DeepSeek and others through one account. It has an endpoint in the same format as Anthropic's, so the change is a web address and a key, not a rewrite.
It does not make Claude cheaper: the per-token prices are the same as buying direct. There is no markup per token, but buying credits by card costs 5.5% (minimum $0.80). Worth it when you want the cheaper non-Claude models. Not worth it just to run Claude.
How I would keep it down
- One API key per job, so the bill tells you what spends
- Move the background jobs to cheaper models first. They have no personality to lose
- Keep the character on the good model, and switch it last, after a side-by-side test
- Cache the instructions. They are the same on every message, and a cache read is a tenth of the price
- Set a daily cap on voice. Gerry has one, and falls back to a free local model when it runs out
The face: a one-off cost
Gerry's idle loop, the 44 seconds of him fidgeting in his chair, is six short video clips made from one image. The whole thing cost $3.39 to render, including a test on a second model. Each extra 10-second behaviour is about $0.70. You pay once, then it plays all day.
Everything else
Airtable and Vercel are free at this size. Claude Code, which writes the app for you, needs a paid Claude plan. That is the whole stack.
Step 1 · Free
Give it a character →