A working coding agent costs between nothing and $200 a month. The plan you already pay for probably covers it, and the expensive part is the time you spend reviewing what it wrote.
| Setup | Per month | What you get |
|---|---|---|
| Claude Pro | $20, or $17 annual | Claude Code and chat, sharing one usage limit |
| Claude Max | $100, or $200 for 20x | The same, with five or twenty times the allowance |
| ChatGPT Free | $0 | Codex included, with the free tier's limits |
| Gemini CLI | $0 | 1,000 requests a day with a personal Google account |
The subscription is the cheap path
Anthropic includes Claude Code in every paid plan and shares the plan's limits with it, so a long terminal session spends the same allowance as a long conversation. OpenAI includes Codex in every ChatGPT plan, Free included.
For one developer working a normal week, Pro and Plus cover it. You find out otherwise by hitting the limit, not by forecasting.
The API path has no ceiling
Pay per token and the bill is whatever you asked for. A ten-step agent loop re-sends the conversation each step: on a 200,000-token history at Opus 5.5's $4 per million input tokens, that is $8 of input for one task before the model writes anything.
Ten tasks a day at that size is $2,400 a month. Nobody does this by accident twice.
The cost that does not appear on the invoice
Reviewing generated code takes real time, and a bad change costs more than the tokens saved. Budget an hour a day for a single agent, and treat any claim that it needs less as unproven until you have run it for a month.
Start with the plan. Move to an API key when you can name the volume.