What running Claude Code in the cloud really costs

Claude Code itself is free to install. What you pay for is the model it talks to and the computer it runs on. Prices below are as of October 2026; the sources are at the end.

Bill 1: the model

A Claude subscription

Plan	Price	Claude Code
Pro	$20 a month ($17 a month billed yearly)	Included
Max	From $100 a month, for 5x or 20x Pro’s usage	Included
Team, standard seat	$25 per seat a month ($20 billed yearly)	Included
Team, premium seat	$125 per seat a month ($100 billed yearly), 5x a standard seat’s usage	Included

Claude Code shares your plan’s usage limits with the rest of Claude, wherever it runs: your laptop, a server, or Anthropic’s cloud sessions. Running several sessions in parallel uses the limits up faster.

The API, paid per token

Model	Input	Output	Cached input (reads)
Claude Opus 5.5	$4 per million tokens	$20 per million	$0.20 per million
Claude Sonnet 5.5	$2 per million	$10 per million	$0.10 per million
Claude Haiku 5.5 (prompts up to 100K tokens)	$0.10 per million	$0.50 per million	$0.01 per million

Why agent sessions get expensive at API rates

A coding agent re-reads its whole context, the files, the conversation, the tool output, on every step. In one public report, a 58-minute Claude Code cloud session on Opus 5.5 produced only 3.4K tokens of output, but read 109.8 million cached tokens, and showed $34.38 at API rates (Reddit). Caching made those reads cheap per token; there were simply a lot of them.

That’s why most people who run Claude Code daily use a subscription: the cost is fixed, and heavy days don’t produce a surprise bill.

Bill 2: the machine

Where Claude Code runs decides whether it keeps working when you close your laptop.

Where it runs	Monthly cost	What you get
Your laptop	$0	Stops when the laptop sleeps
Claude Code cloud sessions	$0 extra	Anthropic’s VM, included in your plan; reclaimed after a period of inactivity
Small VPS (Hetzner CX23, EU)	From €5.49, plus €0.50 for IPv4	2 vCPUs, 4 GB; you set it up and maintain it, and cheap plans are often unavailable
Cube	From $12 running, $4 paused	2 vCPUs, 4 GB, 30 GB SSD, ready in minutes with your repos and agents

The same 4 vCPU, 8 GB machine, priced for a full month (730 hours)

Provider	Per month	Notes
boat.dev	$26.28	4 CPUs, 50 GB SSD, EU; billed per second while running
exe.dev	$35	A 4-vCPU pool shared across your VMs, 200 GB disk
DigitalOcean	$48	Basic Droplet, 4 shared vCPUs, 160 GB SSD
AWS EC2	$111.69	c6a.xlarge on demand; disk and IP extra
Docker Cloud Sandboxes	about $204	Medium size at $0.28 an hour; sessions last up to 24 hours

On Cube, each vCPU costs $6 a month running, with RAM and SSD included, so 4 vCPUs with 8 GB and 60 GB comes to four times that. A machine that pauses when you don’t need it costs less than these always-on figures.

Three typical setups

Setup	Model bill	Machine bill	Good for
Pro + cloud sessions	$20	$0	Occasional tasks you start from the browser or phone
Pro or Max + a VPS you run	$20–$200	A few euros to ~$50	People who like running servers
Pro or Max + Cube	$20–$200	From $12	Daily agents that keep running, with no server to maintain

Six ways to spend less

1. Use a subscription for daily work, and the API only for automation and light use.
2. Match the model to the task. Sonnet 5.5 costs half of Opus 5.5 per token; save Opus for the hard problems.
3. Keep context lean. Every step re-reads it. Use /compact or /clear between unrelated tasks.
4. Don’t run more parallel sessions than you can review. Each one draws on the same usage limits.
5. Pause machines you aren’t using. On Cube, a paused computer costs $2 per vCPU a month and keeps its files.
6. Price the whole month, not the hour. Per-hour prices look small until a machine runs all 730 hours.