Claude Code pricing 2026: Pro, Max, or API key — web and Vietnam App Store prices, session cost, and top-up with a Visa/Mastercard card

Downloading and installing Claude Code costs nothing; the cost comes from each time it calls a Claude model, paid by either your subscription or your API key. Summary before the details:

• Subscription plan. Claude Pro $20/month ($17/month when billed annually, equivalent to a single upfront payment of $200), Claude Max $100 or $200/month. These are US web prices before tax. When purchased through the Vietnam App Store, Pro costs 599,000 ₫ per month. The Free plan cannot run Claude Code.
• API key. No monthly fee; you pay for what you use. Through Kunavo, Claude Sonnet 5 costs $1.40 per million input tokens and $7.00 per million output tokens; an agent step of about 25,000 input tokens and 1,200 output tokens costs approximately $0.0434, while with a warm cache it costs only about $0.020.
• Which should you choose? $20 of Pro buys approximately 460 such steps without a cache. If you use Claude Code almost every day and the Pro limit is still sufficient, Pro is better value; if you code intermittently, the key is better value because months without coding cost nothing.
• Payment. Claude plans purchased on the web accept only cards. Kunavo balances can be topped up from $10 using a card (Visa, Mastercard), Link, Apple Pay, or Google Pay; MoMo, ZaloPay, and VNPay are not available.

Figures checked on October 3, 2026: plan prices from claude.com/pricing, Anthropic support articles about the Pro plan, Max plan, and billing, as well as Claude's page on the Vietnam App Store. Kunavo prices are from the model catalog, updated October 10, 2026. The article uses USD because both Anthropic (on the web) and Kunavo list prices in USD; the only VND figure here is the price Apple sets itself.

Claude plan pricing: on the web (USD) and in the Vietnam App Store (VND)

To use Claude Code with a Claude account, you need a Pro, Max, Team, or Enterprise plan, or a Console account (setup documentation). Vietnam is included in the list of supported countries for both the API and Claude.ai. The web column shows prices on claude.com/pricing and in the support article about the Max plan; the App Store column shows the purchase price in the Claude app with a Vietnamese Apple account. All figures were checked on October 3, 2026.

Three things to know when reading this table:

• Taxes. claude.com/pricing explicitly says that prices exclude tax. The support article about the Pro plan says monthly prices vary by region, are shown in local currency in some places, and may include tax in the price or have it added at checkout, depending on the location. The App Store page does not say whether the VND price includes tax, so this article makes no claim either way.
• Web or app. Anthropic says that the Max price is the web subscription price; purchasing through a mobile app may differ. We do not convert USD to VND ourselves because the actual amount charged to your card also depends on the exchange rate, your bank's foreign-currency fees, and the taxes applied by the checkout page. Compare 599,000 ₫ with the final total shown by the claude.ai checkout page.
• Team seats. Some Vietnamese articles say that Team Standard seats do not include Claude Code. This information is outdated: the support article about Claude Code on Team and Enterprise plans says Claude Code is available on all Team seats (checked on October 3, 2026).

Can plans be purchased with MoMo, ZaloPay, or VNPay?

No, if you buy on the web. The Anthropic billing FAQ says plans purchased on the web can be paid for only by credit or debit card. For in-app purchases, the App Store or Google Play processes the payment using the methods allowed for your account. Vietnamese digital wallets are also unavailable at Kunavo; supported top-up methods are listed in the top-up section below.

Is Claude Code free?

Installation is free; running it costs money. The CLI and extensions for VS Code and JetBrains are free; what is charged is each model call, including calls made by subagents. The Free plan on claude.ai does not enable Claude Code, while the API charges for tokens from the first request, so there is no way to run the model free of charge indefinitely.

If you only want to try it for a few weeks, a prepaid key is the option with the fewest commitments. When you top up Anthropic's API directly, credits expire after one year and are non-refundable (Anthropic Help Center, checked on October 3, 2026). At Kunavo, the minimum top-up is $10, the balance never expires, and there is no monthly fee. That amount covers about 230 agent steps on Claude Sonnet 5, or around 12 tasks of 20 steps each, excluding savings from caching.

Pro/Max 5-hour and weekly limits

A subscription is sold by usage limits, not by tokens: one session limit resets every 5 hours, and one weekly limit applies across all models (Pro plan support article). Claude chat and Claude Code draw from the same pool, so a long chat session also reduces what's left for Claude Code (support article about Claude Code on Pro and Max plans). Anthropic does not publish how many tokens each limit represents; the Claude limits guide (in English) explains the time windows in more detail.

According to Claude Code's cost documentation, these two limits are shared across all models, so switching models with /model will not let you continue; switching only helps when the message names a specific limit (for example, the Opus limit). When blocked, you have four options: wait, upgrade your plan, enable usage credits, or set an API key. Usage credits are prepaid and charged at standard API prices (Anthropic Help Center), and while using them, the cache lasts 5 minutes instead of 1 hour when you are within your plan limits (cost documentation; both checked on October 3, 2026). In other words, usage beyond the limit is charged at API prices either way; paying with a Kunavo key makes Claude models 30% cheaper than Anthropic's price list.

Using an API key: Claude Code is charged differently

Plans and keys do not stack. According to the LLM gateway documentation and the support article about Claude Code on Pro and Max plans (checked on October 3, 2026):

• Having ANTHROPIC_AUTH_TOKEN (or ANTHROPIC_API_KEY, after you approve the key once in interactive mode) makes Claude Code use the key and bypass the plan.
• The 5-hour and weekly limits no longer apply; the key owner pays for the tokens, and the only limit is the balance.
• Your claude.ai login session remains saved but inactive. Remove the variable and Claude Code returns to your plan (gateway connection documentation); /status shows which billing method is active.

With a key, everything is measured in tokens. File contents, command output, and conversation history are input; code edits, explanations, and thinking tokens are output. The table below shows prices per million tokens (input / output), with Anthropic prices from claude.com/pricing (checked on October 3, 2026) and Kunavo prices from the catalog (updated October 10, 2026):

The “one step” column uses one common unit for every language version of this guide: one Claude Code request to the model while working on a task, consisting of approximately 25,000 input tokens (system prompt, history, and recently read file content) and 1,200 output tokens (a code edit or an explanation), calculated as if there were no cache.

Prompt cache: most of the cost is here

Claude Code does not “remember” between requests; each step sends the entire conversation again. The repeated portion is much cheaper through prompt caching than regular input, so the actual cost is usually lower than the “one step” column above. Kunavo cache prices (per million tokens, 5-minute cache, multiplier applied to each model's input price):

The catch is the cache lifetime: 5 minutes when using an API key, 1 hour with a plan (cost documentation, checked on October 3, 2026). Go out for lunch and come back, and the first message has to load the entire context again at the cache-write price instead of the cache-read price. For 200,000 context tokens on Claude Sonnet 5, that reload costs about $0.35 at Kunavo, compared with $0.028 if the cache is still warm. If you're returning to a different task, enter /clear first so you don't pay for the old context. For details about caching, see the caching documentation (in English).

Calculate a Claude Code session

In the cost documentation (checked on October 3, 2026), Anthropic provides a sample /usage result: 1,200 input tokens, 5,300 output tokens, 940,000 cache-read tokens, and 50,000 cache-write tokens, running on Sonnet 4.6, for a total of $0.55 at list prices. The table below keeps those four figures unchanged and only substitutes the unit prices for four other models. Cache writes are calculated at the 5-minute cache price (the default when using a key); differences in tokenizers between models are ignored; thinking tokens are included in output; background calls, which the documentation says are usually under $0.04 per session, are excluded.

With Claude Sonnet 5, the sample session costs about $0.258 at Kunavo ($0.368 according to Anthropic pricing). Nearly 51% of that is cache-read cost, meaning this portion becomes larger as the session gets longer. $20, exactly one month of Pro, pays for approximately 77 such sessions. Switching to Claude Opus 5.5 costs about $0.384, Claude Opus 5 costs about $0.645, and Claude Haiku 4.5 costs about $0.129.

The figures above are illustrative calculations, not measurements or limits. To calculate your own session, enter /usage to get the four token figures, then put them into this Python snippet:

# Chi phí một phiên Claude Code từ chính số token của nó (USD, giá Kunavo).
# Lệnh /usage trong Claude Code hiển thị bốn con số của phiên hiện tại.
BANG_GIA = {  # cho 1 triệu token: (đầu vào, đầu ra, đọc cache, ghi cache)
    "claude-sonnet-5": (1.40, 7.00, 0.14, 1.75),
    "claude-opus-5-5": (2.80, 14.00, 0.14, 3.50),
    "claude-opus-5": (3.50, 17.50, 0.35, 4.375),
    "claude-haiku-4-5": (0.70, 3.50, 0.07, 0.875),
}

def chi_phi_phien(mo_hinh, dau_vao, dau_ra, doc_cache, ghi_cache):
    i, o, r, w = BANG_GIA[mo_hinh]
    return (dau_vao * i + dau_ra * o + doc_cache * r + ghi_cache * w) / 1e6

# Phiên mẫu trong tài liệu chi phí của Anthropic:
print(round(chi_phi_phien("claude-sonnet-5", 1_200, 5_300, 940_000, 50_000), 3))  # 0.258

Note: the USD amount shown by /usage is an estimate calculated by Claude Code on your machine using Anthropic's pricing. When running through Kunavo, the actual amount charged appears in the usage history in the dashboard. Other scenarios can be calculated with the Claude token cost calculator. As a benchmark: the cost documentation says that at organizations deploying Claude Code, the average developer spends about $13 per day of use, $150–250 per month, and 90% of users spend under $30 per day of use — Anthropic averages, not Kunavo pricing.

Plan or key: calculate based on how you code

No need to guess. Estimate the number of steps per month, multiply by the price per step, and compare it with the plan fee. The table below does this for three usage patterns on Claude Sonnet 5 at Kunavo: the “cold cache” column is the worst case ($0.0434 per step), while the “warm cache” column is when 20,000 of the 25,000 input tokens are read from cache and the rest are written to cache ($0.020 per step).

Reversing the calculation gives us the break-even threshold: plan fee ÷ price per step = the number of steps per month from which the plan is cheaper. Using Anthropic's prices before tax and the price per step without caching, Pro breaks even at approximately 460 steps, Max 5x at approximately 2,300 steps, and Max 20x at approximately 4,610 steps. With a warm cache, Pro's threshold rises to approximately 1,000 steps. One unknown remains: Anthropic does not publish how many steps each plan's limit supports, so “the plan is cheaper” is true only if the limit lets you reach that level.

Keys remain useful for work that subscriptions do not handle neatly: running in CI, automating scripts, or assigning separate budgets to team members with individual keys. The two approaches also do not exclude each other over time: keep Pro for normal days, and use a key for the hours blocked by limits instead of buying usage credits at API prices.

Should you upgrade to Max?

The difference between Max 5x and Pro is $80 per month. That same $80, added to Kunavo, runs approximately 1,840 steps on Claude Sonnet 5 even with a cold cache—about 37 afternoons of work at 50 steps per afternoon, with each afternoon costing approximately $2.17. If Pro only blocks you for a few afternoons each month, paying with a key for those afternoons is cheaper than upgrading. Max is worth it when you hit the Pro limit often enough that the blocked work, calculated by key, exceeds $80.

Configuring Claude Code to use a Kunavo key

Claude Code reads the ANTHROPIC_BASE_URL variable, so pointing it to Kunavo is sufficient; no plugin is needed. Kunavo communicates with Claude Code through the Anthropic Messages API. If Claude Code is not installed yet, follow the Claude Code installation guide first. On macOS or Linux, add the following lines to ~/.zshrc (or enter them directly in the terminal if you are only testing for one session):

export ANTHROPIC_BASE_URL=https://api.kunavo.com   # chỉ địa chỉ, không thêm /v1
export ANTHROPIC_AUTH_TOKEN=sk-kn-...
export ANTHROPIC_MODEL=claude-sonnet-5
export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-5-5
export ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-5
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5

On Windows, open PowerShell:

# Chỉ có hiệu lực trong cửa sổ PowerShell hiện tại
$env:ANTHROPIC_BASE_URL = "https://api.kunavo.com"
$env:ANTHROPIC_AUTH_TOKEN = "sk-kn-..."
$env:ANTHROPIC_MODEL = "claude-sonnet-5"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = "claude-opus-5-5"
$env:ANTHROPIC_DEFAULT_SONNET_MODEL = "claude-sonnet-5"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL = "claude-haiku-4-5"
claude

Three common mistakes. First, ANTHROPIC_BASE_URL is only the server address; don't append /v1, because Claude Code adds /v1/messages itself. Second, use ANTHROPIC_AUTH_TOKEN (which sends the key in the Authorization: Bearer header) instead of ANTHROPIC_API_KEY (which sends it in x-api-key) (gateway connection documentation, checked on October 3, 2026). Kunavo reads the key from either header; the difference is that ANTHROPIC_AUTH_TOKEN takes effect immediately, while ANTHROPIC_API_KEY requires your one-time approval in interactive mode. Third, for persistent configuration, put the settings in the env block of the user settings file; if a variable is present in both the shell and this file, the file value takes precedence:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.kunavo.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-kn-...",
    "ANTHROPIC_MODEL": "claude-sonnet-5",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5-5",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5"
  }
}

Put the key only in the user settings file; do not put it in the project's .claude/settings.json because that file is committed to the repository. When using the VS Code extension, set the variable in claudeCode.environmentVariables in VS Code's user settings.

Why the model names need to be specified. According to the model configuration documentation (checked on October 3, 2026), when using the Anthropic API, Claude Code runs Opus 5.5 by default, the opus alias is Opus 5.5, and the sonnet alias is Sonnet 5.5. Kunavo does not offer Sonnet 5.5: without ANTHROPIC_DEFAULT_SONNET_MODEL, /model sonnet, the execution phase of opusplan, and every subagent with model: sonnet all call a Kunavo model that is not served and receive a 404 error. The configuration above starts with Claude Sonnet 5 to reduce cost, pins the sonnet alias (the ANTHROPIC_DEFAULT_SONNET_MODEL variable) to Claude Sonnet 5, lets ANTHROPIC_DEFAULT_HAIKU_MODEL handle background tasks too, and pins the opus alias (the ANTHROPIC_DEFAULT_OPUS_MODEL variable) to Claude Opus 5.5 ($2.80 / $14.00). Claude Opus 5.5 requires Claude Code v2.1.280 or later (run claude update first on older versions), always enables thinking, and thinking tokens are billed as output (cost documentation).

Check: run claude and enter /status. The Anthropic base URL line must be https://api.kunavo.com, and the Auth token line must point only to ANTHROPIC_AUTH_TOKEN; if either is missing, the variable did not reach the session. Create a sk-kn- key on the key management page after signing up. Full integration guide (English): Claude Code with Kunavo; the English version of this pricing article: Claude Code pricing.

Top up from Vietnam: Visa/Mastercard cards, Link, Apple Pay, Google Pay

Kunavo is an API gateway compatible with both Anthropic's and OpenAI's interfaces, and the money you add is a prepaid balance for your API key. Top-ups go through Stripe's checkout page. Prices are listed in USD, but Stripe may convert them to VND on the page before you click to pay. According to Stripe's Adaptive Pricing documentation (Vietnam is on the list of supported markets; checked on October 3, 2026), the displayed exchange rate includes a 2–4% conversion fee paid by the buyer. Choose to pay in USD to avoid this fee, but your card issuer may apply its own exchange rate and foreign transaction fees.

• Minimum top-up $10. Larger top-ups receive an additional bonus: top up $100 to get $110 in the balance; top up $1.000 to get $1.200 in the balance; top up $5.000 to get $6.250 in the balance.
• Prepaid balance, no monthly fee, no expiration; failed requests are not charged.
• Automatic reload works only with a saved card or Link.
• Kunavo does not issue VAT invoices, including electronic invoices; view top-up history (date, amount, status) on the Billing page.
• This balance is only for API keys. It cannot pay for Claude Pro or Max: those subscriptions are purchased directly from Anthropic on the web, or through the App Store / Google Play.

Get started: create a Kunavo account, top up your balance, and create a key. Prices for all models are on the Kunavo pricing table. Detailed guides to paying with mobile wallets (English): Apple Pay and Google Pay.

When a key cannot replace a subscription

Claude Code's agents, tools, subagents, and MCP servers all run on the Messages API, so they work normally with a gateway key. Two features do not: Remote Control and voice input require a claude.ai identity, so they are disabled when ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, or an apiKeyHelper is active (gateway connection documentation, checked on October 3, 2026). Anthropic also states that it does not validate third-party gateways (LLM gateway documentation).

In short: if you code every day and the Pro limit is still sufficient, keep Pro. A key is useful when you code in bursts, need a separate budget for each person, run CI, or only need a few extra hours after the subscription limit has been reached.

Frequently asked questions