Claude Code Max Plan 2026 | US$100/200 Monthly, Pro US$20
Claude Code plans: Pro US$20/month, Max 5x/20x US$100/200, or pay-as-you-go API. Covers usage limits, tips and Taiwan expense claims (checked 2026-09).
Plan and documentation check date: 2026-09-17 (Taipei time). This article merges the former "Claude Code Fees" and "Claude Code Tutorial" articles and keeps the
/blog/claude-code-feeURL. Prices are US dollar list prices from the official websites and exclude applicable taxes; mobile storefronts and actual checkout amounts may differ. This update checked the official documentation only. It did not re-run the CLI installation or any paid workloads.
How much does Claude Code cost per month? The short answer
First decide what you need: a terminal tool, team seats, or a programmatic API.
| Your situation | Which path to evaluate first | Official list price or billing method |
|---|---|---|
| An individual writing code in the terminal, no usage history yet | Pro monthly; observe your actual limits first | US$20/month; US$200 per year |
| Pro often runs into limits during work | Max 5x / 20x | US$100 / US$200 per month |
| A company that wants centralized account management | Team / Enterprise | Depends on seat type, usage and contract; do not apply the individual Max price |
| Putting models into a product or automated workflow | API | Billed by actual billable usage such as tokens, not the Pro monthly fee |
For Pro pricing, see the Claude official pricing page. For the two Max tiers, see the official Max documentation.
Without usage records, I start with monthly billing and do not rush into annual payment or Max. This does not mean Pro is enough for everyone. The model, project size, long tasks and parallel sessions all affect limits. Confirm the tool suits your work before considering an upgrade.
Company purchasing adds one more question: are you buying an official subscription seat, or an API service from a local platform? Invoices and claim documents must match the real transaction. Two things both being called Claude does not make them interchangeable.
Claude Code, Claude App and the API: billing is not the same thing
Claude Code is a coding agent that can read files, modify code and run tools in a project. Claude App is a different interface. The sign-in methods that Claude Code currently lists in the official documentation include Pro/Max, Team/Enterprise, Console, and supported cloud deployment options. See the official authentication documentation.
The free Claude chat plan cannot be used directly as a Claude Code subscription. You also should not calculate your API cost per call by dividing the Pro monthly fee by the number of chat messages.
A single Claude Code task may include multiple model calls, file reads and tool results. "Please fix this bug" looks like one sentence, but the billed usage is much more than that sentence. To compare costs, compare the total usage and time needed to complete the same task.
Pro, Max 5x and Max 20x: how to choose
| Individual plan | List price | Conditions to note |
|---|---|---|
| Pro | US$20/month, or US$200/year | An annual plan is paid up front for the full year, not about US$17 deducted each month |
| Max 5x | US$100/month | Higher usage than Pro, but not unlimited |
| Max 20x | US$200/month | Higher usage, still subject to plan limits |
Pro's annual plan costs US$40 less than twelve monthly payments of US$240. However, testing the workflow with the monthly plan first avoids buying a full year and then finding it unsuitable. Max is currently billed monthly, and web and mobile prices may differ.
Identify which limit you hit before deciding to upgrade
The official Max documentation lists a five-hour session usage window and also weekly limits; other model or feature limits may have separate rules. So "waiting five hours will always let you continue" is not a valid assumption. Check the actual reset time shown for your account.
Claude Code's /usage command shows your usage and plan status. The subscription usage bar and the current session's dollar estimate are not the same thing as a bill. The official cost documentation also says API charges should be based on actual billing in the Console. See the usage and cost documentation.
Keep a one-week record. It is more useful than saying from memory that you "often run out":
| Item to record | Why it matters |
|---|---|
| Whether you hit the five-hour limit, the weekly limit, or another spending cap | An upgrade may not solve the same kind of limit |
| The model, task and context length at the time | Shows whether one long task consumed most of the quota |
| The time your work was actually interrupted | Compare it with the upgrade difference, rather than counting all waiting time as a loss |
| Whether separately paid usage credits were enabled | Avoids mistaking "work can continue" for "included in the monthly fee" |
Upgrading from Pro to Max 5x is a US$80/month list price difference; from 5x to 20x is another US$100. Multiply the actual time affected by your time cost and compare that with the difference. Occasionally hitting a limit does not necessarily justify paying the higher monthly fee.
Team and Enterprise: do not estimate from the individual plan
Company needs also include members joining and leaving, permissions, budgets, data policies and procurement documents. The current official price list separates Team's standard and premium seats. Enterprise may also combine a seat fee with API rate-based usage. Do not describe it as "a fixed monthly fee per person, with the same quota within the allowance."
When purchasing, first get the formal terms for the plan you choose: which members can use Claude Code, whether there are extra usage fees, who can set spending limits, and which account the data policy applies to. Use the official Team / Enterprise price list and written quotes as the reference, and do not use the individual Pro price to budget for a team.
How API pay-as-you-go billing works
Four variables affect cost: input tokens, output tokens, caching conditions, and the model. Input is not only the new question. It may also include instructions, the message history sent, files and tool results.
You can check the current customer price for each model in the BazaarLink model catalog; for direct official purchases, check the official price list. Keep the two price sources separate. Do not use official costs to calculate a BazaarLink bill.
A recalculable example
Assume 50 API requests per day, an average of 2,000 actually billed input tokens and 500 output tokens per request, for 30 days:
Monthly input 50 × 2,000 × 30 = 3 million tokens
Monthly output 50 × 500 × 30 = 0.75 million tokens
Monthly cost US$ = 3 × input price per million + 0.75 × output price per million
This is only an example with stated conditions, not an average for users. If one task calls the model ten times in sequence, count ten requests. Reasoning, caching or other billable items should be handled separately according to actual usage and channel rules.
The crossover point between subscription and API cannot be set by a fixed multiple
"The same work uses five to ten times more tokens through the API" is not a controlled measurement result in this article, so it cannot be used to estimate your bill. The sign-in path, model, tool workflow, context and usage limits all change the result.
The comparison method is to choose a few tasks you do often and record for each the completion status, manual retries, total usage and limits. If you work in Claude Code every day and the current subscription limits meet your needs, there is no reason to switch just because of the word "API." When you need to embed in a product, call per project, or manage a multi-model workflow, then evaluate the API.
Claude Code, Codex and Cursor should not be compared by monthly fee alone, because their interfaces, plan inclusions and separately billed usage differ. For comparison, see the Codex fees and Cursor fees articles. Do not treat "only Claude Code can use the API" as a reason to choose a tool.
How to reduce unnecessary usage: five actions
Clear context after finishing unrelated tasks
/clear suits switching to unrelated work. For a long task that still needs earlier conclusions, consider /compact. Do not clear after every small step. Otherwise you resend background and reread files, which can increase usage.
Pair tasks with /model
Simple fixes and tests do not necessarily need the most expensive model; for complex design you can evaluate others separately. Whether a model can complete your kind of work should be judged by actual testing, not by assuming "Sonnet is always enough" or "Opus is always best."
Write rules you keep repeating into CLAUDE.md
Run /init to generate a first draft, then keep only the instructions that truly affect your work. For example, the following is a draft; replace the commands with the versions that actually exist in your project:
# Project rules
## Commands
- Test: pnpm test -- --run
- Type check: pnpm typecheck
## Rules
- Add matching tests after changing the API
- Confirm the purpose before adding a package
- Do not refactor src/legacy/ in passing; it has external dependencies
CLAUDE.md itself is also loaded into context, and longer is not always better. Put the rules you need repeatedly into it, and store large background material separately to read when working.
Review the plan before approving large changes
A plan can rule out wrong directions, but there is no evidence that supports "the cost will definitely drop to one tenth." Give explicit acceptance criteria, then check whether the plan actually covers them.
Make instructions reproducible
"Improve the codebase a bit" makes the tool guess the scope. Give file paths, error messages and test commands, for example:
First run pnpm test -- --run src/cart/total.test.ts and confirm the third case fails.
Explain the cause and the smallest change scope, then rerun the same test after the change. Do not touch unrelated modules.
This is an example of a workflow, not the result of tests actually run for this article. For more cost-control details, see the official cost documentation.
Installation and day one: run one small task
The official documentation currently recommends the native installer; Windows PowerShell and macOS/Linux use different methods. Go to the official installation documentation and choose the method that fits your system and shell, and confirm the source before running any remote install script.
If you choose npm installation, the current documentation says Node.js 22 or later. This is a requirement of the npm installation path, not something every installation method needs.
node -v
npm install -g @anthropic-ai/claude-code
claude --version
Run claude in your project directory, then sign in with the official account or deployment method you selected. Keep your first task small: for example, ask it to explain the data flow of one module, or reproduce a known test failure. Understanding the code does not mean later changes are necessarily safe; always review the diff yourself.
On day one, learn these five commands first
| Command | Purpose |
|---|---|
/init | Generate a first draft of the project CLAUDE.md; needs manual review |
/clear | Clear the current conversation context when switching to unrelated work |
/compact | Compress the current conversation; check that important conclusions were kept |
/model | Choose a model available for your account and path |
/usage | View usage and plan status; estimates do not replace the actual bill |
Non-interactive mode can go into scripts, for example:
claude -p "Summarize the unreleased items in CHANGELOG.md and output release notes" > release-notes.txt
This command calls the model and writes a local file. When used in CI, you must also set permissions and a budget. Do not treat it as a text tool with no cost.
Permissions and recovery: Git only protects the files it manages
Putting commonly needed tests, type checks and read-only Git commands on an allowlist can reduce repeated approvals. Pushes, deployments, database writes, deletions, payments and messages should be evaluated separately. Do not approve everything just because the tests pass.
Before a large change, check git status and save the work you have completed. After the change, review the diff yourself and rerun the tests. Do not treat git checkout . as a general safety net. It discards unsaved changes to tracked files, cannot recover sent emails, deployments, database changes or leaked keys, and may not cover untracked files. Control permissions and external side effects before acting.
Taiwan companies filing expenses: separate subscriptions, API and claim documents first
If you bought an official Anthropic subscription, confirm payment and documents with that counterparty or purchase channel first. BazaarLink cannot open a new invoice from this site for a separate subscription you already bought from the vendor.
The BazaarLink API service is a separate transaction. Depending on the actual service and checkout process, NT dollar payments, unified invoices and the company's tax ID number can be handled. This does not mean that buying the API gives you an official App or Claude Code subscription seat, nor does it guarantee that any third-party endpoint supports every feature of the official tools.
In addition, whether an overseas receipt can serve as an expense voucher and whether you get a unified invoice that allows input tax deduction are not the same question. Do not translate "no Taiwan unified invoice" directly into "the company absolutely cannot claim it." Your accountant needs to confirm this based on the transaction, payment and applicable rules. For details, see the AI API Taiwan expense guide and Claude tax ID and transaction documents.
A grant deadline does not mean you can only buy API top-ups
When a school, research project or public-sector body has a claim deadline, first confirm the purchasable items, service period, payment date and voucher date. Monthly subscriptions, annual subscriptions and prepaid API credit may each have different conditions for claims. Prepaid API credit also does not automatically meet grant rules just because it was paid all at once.
Before purchasing, you can prepare a table like this:
| Item to confirm | Who to get it from |
|---|---|
| Service content, period and account user | Written description from the supplier |
| NT dollar quote, payment deadline and document type | The supplier's quote and checkout terms |
| Keys, models and spending management method | Internal technical owner |
| Data retention, routing and usage policy | The service's current privacy/security documentation, not inferred from advertising copy |
| Whether the expense category is allowed and when the claim must be completed | Company accountant or project administrator |
Without written evidence, do not treat "no content retention and no training" as a purchase commitment that holds for every model and channel.
When is BazaarLink suitable?
If you mainly work in Claude Code every day and the official subscription's features and limits already meet your needs, keep using it. If you need programmatic calls, multi-model workflows, individual API keys and budget management, or NT dollar payments and documents for the BazaarLink API, then evaluate BazaarLink.
Buy the right service first, then compare unit prices. Official subscription seats, API credit and invoice supplementation are three separate questions. Separating them usually saves more rework than buying the highest plan directly.
FAQ
How much does Claude Code cost per month?
As of an official pricing check on 2026-09-17: Pro is US$20 per month or US$200 per year; Max 5x is US$100 per month; Max 20x is US$200 per month. Team and Enterprise depend on seat count and plan terms, and API is billed by usage. List prices exclude applicable taxes, and mobile storefronts and actual checkout amounts may differ.
Will Claude Pro always be enough?
Not guaranteed. The model, project size, context, long tasks and multiple sessions all affect usage. Check your actual limits with /usage and record for a week how often work is interrupted before deciding whether the Max upgrade difference is worth paying. A monthly fee is not unlimited usage.
If I hit the Claude Code limit, can I always use it again after five hours?
Not necessarily. The official Max documentation describes a five-hour session window and also weekly limits. Look at the limit you actually hit and the reset time shown for your account. Enabling usage credits may also create usage charges beyond the monthly fee.
Is a Claude subscription or the API cheaper?
This article does not establish any fixed 5x-to-10x conversion. Use the same task, model and workflow, and record completion, retries, usage and limits, then compare your own costs. If an official subscription covers your terminal work, there is no need to switch to the API just for that. Evaluate the API for products, automation or multi-model workflows.
Can a Taiwan platform issue an invoice for a Claude subscription I already bought?
BazaarLink cannot issue an invoice from this site for a separate subscription transaction you already made directly with the vendor. This site's API is a different service and transaction, and it is not the same as an official subscription seat. Whether overseas receipts can be used for claims and whether input tax can be deducted also differ. Please confirm with the actual counterparty of the transaction and your company accountant.
If a research grant has a claim deadline, can I only buy API top-ups?
No. For monthly or annual subscriptions, or prepaid API credit, confirm whether they are allowed based on the expense category, service period, payment date and voucher date. A one-time top-up does not automatically meet claim rules. Ask your project administrator or accountant first.
Can Git restore everything Claude Code has done?
No. Git can only help manage the files it tracks. It cannot undo external side effects such as deployments, database writes, emails, payments or leaked keys. Do not treat git checkout . as a general safety net, because it also discards unsaved changes to tracked files. Save your existing work first and control permissions.
TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API