Claude Code pricing in 2026: what every plan actually costs

Every way to pay, priced

Anthropic sells access to Claude Code five different ways, and most write-ups treat them as one ladder. They are not one ladder. Three are prepaid subscriptions with usage windows, one is metered with no windows at all, and one is a top-up that bridges the two.

Where the money actually goes on a team plan

On Team and Enterprise, each member's Claude Code usage draws from a per-seat allowance on a rolling five-hour window and a weekly window, and that allowance is shared with Claude chat and Cowork. So a developer who spends the morning in chat has less left for agents that afternoon. Budget a coding seat higher than a chat seat: every Claude Code turn carries file contents, tool results, and multi-step reasoning.

What a subscription actually buys

Subscription usage is not sold in dollars or in tokens you can see. It is a rolling five-hour window plus a weekly cap, and the plans differ by a multiplier on both. That means "how much Claude Code can I run" has no single number: it depends on the model, the size of your context, and how noisy your tool output is.

• Pro handles short, focused sessions comfortably. Three or four half-hour agent tasks a day rarely hit a wall.
• Max 5x is where a normal full working day of agent-assisted coding stops being interrupted, as long as you are not on the largest model for everything.
• Max 20x is what you buy for parallel agents, long autonomous runs, or an Opus-by-default habit.

There is a third limit almost nobody accounts for: a separate Opus allowance. When you exhaust it the CLI says You've hit your Opus limit and you can keep working by running /model and switching. The session and weekly limits are shared across all models, so switching does not help with those two.

API pricing, and where the break-even sits

The API is a different billing model entirely. No monthly fee, no five-hour window, no weekly cap. You pay for every input and output token, and Claude Code will happily spend them.

The three multipliers that decide your bill

Agent sessions are input-heavy in a way chat is not. Every tool result, every file read, every diff goes back into context on the next turn, so a long session re-sends a large prompt many times. Prompt caching absorbs most of that, and the caching multipliers are what make the arithmetic work.

• 5-minute cache write: 1.25x the base input price.
• 1-hour cache write: 2x the base input price.
• Cache read (a hit): 0.1x the base input price. This is the number that saves you.

A 5-minute cache pays for itself after one read; a 1-hour cache pays for itself after two. In practice a long Claude Code session is mostly cache reads, which is why a session showing a million tokens of traffic can still cost less than a dollar.

What it actually costs per developer

Anthropic publishes real numbers rather than making you guess. Across enterprise deployments, the average is around $13 per developer per active day and $150 to $250 per developer per month, with 90% of users staying below $30 per active day.

Where the API still wins

• CI and automation. A GitHub Action cannot hold an interactive subscription session.
• Spiky usage. Two heavy days a month costs less metered than $100 standing.
• No window risk. API calls do not stop because a rolling window filled up. Before a deadline, that predictability has value.
• Multi-tenant products. If you are building on top of Claude, you need API keys regardless.
• Batch work. The Batch API is 50% off input and output, which no subscription matches for offline jobs.

Usage credits: the option most people miss

There is a fifth rail that sits between "prepaid plan" and "metered API": usage credits let you keep working past your plan's limit and pay for the overflow. Run /usage-credits in a session after signing in with a claude.ai subscription; the command is not available under API-key authentication.

The five ways people overpay

1. Defaulting to the largest model for everything

Opus 5 is 5x the output price of Haiku 4.5 and noticeably slower. A large fraction of agent work (running tests, renaming things, applying a described change, writing boilerplate) does not benefit. Dropping routine work to Sonnet or Haiku is the single biggest lever on both cost and rate-limit headroom, and it is one keystroke: /model.

2. Letting context grow unchecked

Because every turn re-sends the conversation, a session carrying 150k tokens pays that on every subsequent message. Anthropic's own guidance is blunt about it: a one-line question in a session that has been open all day still draws usage for the whole conversation. /clear between unrelated tasks costs nothing, while /compact reads the whole conversation in order to summarise it and is itself a large request.

3. Paying for cache misses you could have avoided

Your first message after a break longer than the cache lifetime reprocesses your full context at the full input rate. The lifetime is an hour on a subscription, five minutes on an API key or once you are on usage credits. If you know you are stepping away for two hours, clear the session before you go rather than after you come back.

4. Buying Max 20x for one heavy week a quarter

Subscriptions do not roll over. If your usage is genuinely bursty, Pro plus an API key for overflow beats a standing $200.

5. Paying twice for the same work

Plenty of people hold a Claude Max subscription and a Cursor or Copilot seat and run agents on both. That is fine if each is doing a distinct job. It is waste if the agent work quietly consolidated onto one of them and nobody cancelled the other.

Stacking subscriptions instead of upgrading

People arrive at this independently once they hit limits regularly: rather than buying a bigger single plan, run two accounts. Two Max 5x subscriptions at $100 each cost the same as one Max 20x at $200, and give you two independent five-hour windows instead of one larger one.

Which is better depends on the shape of your work. One Max 20x gives a single session more room. Two Max 5x accounts give you two sessions that cannot starve each other, which is what you want if you run agents in parallel.

The mechanical problem is that the Claude CLI stores one credential per machine user, so a second login clobbers the first. Isolating them means pointing each session at its own config directory via CLAUDE_CONFIG_DIR.

# Give the second account its own config root
export CLAUDE_CONFIG_DIR="$HOME/.claude-work"
claude   # this session logs in and stores creds under ~/.claude-work

# Your original account is untouched under the default ~/.claude
# Check which one you are on before a long run
claude
> /status

Which plan should you buy

Start one tier lower than you think you need. Upgrading takes thirty seconds and the wasted month you avoid is real money. The signal to upgrade is not "I hit a limit once", it is "I hit limits during work I had planned to finish that day", two weeks running.

Questions people ask

Keep reading