Claude Code pricing in 2026: what a one-person business really pays

Subscription plans don't count dollars. They count usage in two windows: a rolling five-hour window and a weekly one. When you hit either, Claude tells you when it resets. Ultra drops the five-hour window for verified accounts and keeps the weekly one.

What fills the window fastest, from Anthropic's Claude Code cost docs:

• Long sessions. Claude Code sends your whole conversation with every request, so a one-line question in a session that has been open all day still costs the whole history.
• Coming back after a break. The first message after the cache expires re-reads your full context.
• Subagents, workflows and scheduled tasks. Each one sends its own requests on top of yours, and a scheduled task fires even while you're away.
• Agent teams. Anthropic puts them at roughly 7× the tokens of a normal session when teammates run in plan mode.
• The model. Opus costs more of your allowance than Sonnet, and Sonnet more than Haiku.

Type /usage inside Claude Code to see where your allowance went. On paid plans it breaks recent usage down by skills, subagents, plugins and connected servers, and suggests changes for anything behind 10% or more of it.

Anthropic publishes its own figure for API-billed Claude Code: across enterprise deployments, about $13 per developer per active day and $150–250 per developer per month, with 90% of users under $30 per active day.

Put that next to the plans. If you work in Claude Code most days, $150–250 a month on the API sits above Max at $100 and around Ultra at $200. On those averages, a subscription is the cheaper way for one person to use Claude Code every day. The API makes sense for occasional use, for a team that wants every dollar metered, or for Claude Code running inside your own automation.

Start one notch lower than you think. Move up when the limit stops real work, not when it shows a warning. Usage credits at API rates cover a rare heavy week without a bigger plan.

That covers building. A system that runs every day has a second bill.

Claude Code is how a tool gets built. Once it runs, the tool calls the Claude API on every record: every reply, every call, every report. That bill follows your volume, not your plan.

An example to size it, with assumptions, not our bill: one record sends 2,000 tokens in (instructions plus the text) and gets 1,000 tokens out (the answer plus thinking). At 15 November 2026 prices:

Pick the smallest model that gets your 20 test records right. The per-item costs we have measured on our builds were for data and images: about £0.18 of data per qualified lead for the broker, and about $0.138 per render for the kitchen visualizer, which runs on Gemini's image model. On the example above, a Haiku call is under a tenth of a cent.

From Anthropic's cost docs, in the order we'd apply them:

1. 1.Clear between unrelated tasks. /clear starts fresh. Name the session first with /rename so you can /resume it later.
2. 2.Match the model to the job. Sonnet for most coding, Opus for hard design calls, Haiku for simple subagents.
3. 3.Keep CLAUDE.md short. It loads at the start of every session. Anthropic suggests under 200 lines; move long, task-specific instructions into skills, which load only when used.
4. 4.Turn off connected servers you aren't using. /mcp lists them; /context shows what fills the window.
5. 5.Plan first on big changes. Plan mode (Shift+Tab) agrees the approach before Claude writes code, which saves expensive rework.
6. 6.Be specific. "Add input checks to the login form in auth.ts" lets Claude read only what it needs; "improve the app" sends it scanning everything.

We add one rule of our own. Each project keeps a short knowledge-base folder with a map file. A session reads the map first, then only the files its task needs, so nobody pays to re-read the whole project every morning.

• The job doesn't need code. A weekly report or an inbox clean-up fits Cowork on the same plan. We compare the three in Claude chat vs Cowork vs Claude Code.
• The rule is fixed. "When the invoice is paid, mark the deal won" needs a one-line automation, not a model and not a build.
• Nobody will own the code. Claude Code writes a working first version fast. Someone still has to review logins, data access and backups, and fix it when an outside service changes. If nobody will, the cheapest build is the one you don't start.
• You're paying for Ultra out of worry. Check /usage for a week first. Anthropic says unexpectedly high spend usually traces back to long sessions that were never cleared, or Opus left as the default model.
• You'd use it twice a month. Pay per token on the API, or stay on Pro.

1. 1.Start on Pro, billed monthly, for the first month.
2. 2.Build one real thing: a small internal tool, a script, a report generator.
3. 3.Run /usage at the end of each day. Note when you hit the five-hour limit and what you were doing.
4. 4.Hit it most days on real work? Move to Max. Running agents in parallel and still blocked? Ultra.
5. 5.If what you built will call Claude on every record, size that bill separately: 20 real records, the smallest model that gets them right, times your monthly volume.

For the bigger picture of what Claude can run in a small business, start at Claude for small business.