Claude Code Costs 2026 — Subscription and API Pay-as-You-Go, Actual Amounts and Break-Even Point

Claude Code has two pricing models, and debates about “how much does it really cost?” often continue because the two get conflated. A subscription has a fixed monthly price and includes usage windows; an API key is billed by token, with no monthly fee or usage limits. This page puts the current figures for both options, the calculation for choosing between them, real monthly costs, and how to pay from Taiwan side by side.

Rates verified on October 5, 2026, using claude.com/pricing and Anthropic's API pricing documentation as sources. Kunavo's per-token rates are read live from the model catalog.

Compare the two options side by side

Subscription plan pricing

Claude Code is included with claude.ai paid plans. Anthropic's installation documentation explicitly says the Free plan can't be used with Claude Code. Pro costs $20/month ($17/month equivalent when billed annually, with $200 paid upfront), covering one developer working within the plan limits. Max starts at $100/month and is for people who hit the limits every day. Team and Enterprise add per-seat allowances and spending controls for administrators. The rules for limits and resets are covered in Claude Pro usage limits. For the equivalent options on OpenAI, see Codex pricing.

From Taiwan, two things directly affect you. First, billing is in dollars and recurring, and requires a credit card enabled for international transactions; your card issuer's foreign transaction fee is charged again every month. Second, usage limits are shared with Claude chat: an afternoon spent discussing things in chat uses up the same window's allowance for coding.

API pricing

There is no monthly fee for this option; you pay by token. The table below shows USD prices per 1M tokens, alongside Anthropic's official rates for easy comparison.

Cached input costs 10% of the input rate, and Claude Code rereads the entire conversation at every step. This means that in real sessions, caching often has a bigger effect on the bill than which model you choose.

Break-even point: cost per agentic step

Choosing between a subscription and API depends on how many agentic steps you run per month—one step is one back-and-forth with the model within a task. In practice, a step is about 25,000 input and 1,200 output tokens (including system prompts, conversation, and file context as input). Using the rates in the table above:

So Pro at $20/month is equivalent to about 460 Claude Sonnet 5 steps, or roughly 23 tasks with 20 steps each. Below that amount, usage-based billing is cheaper, and months without activity cost $0. Above that amount, a subscription is better value if you don't mind hitting the plan limit.

One claim circulating on Reddit is that “a subscription is 36 times cheaper than the API.” This is true only under the assumption that you “fully use your plan allowance,” which is usually left unstated. Using verifiable numbers: 2,000 steps in a month (about 91 per workday) costs $86.80 at Kunavo's Claude Sonnet 5 rates, compared with $20 for Pro. That's 4.3 times. The ratio changes linearly with usage; in low-usage months, it reverses. The real question isn't the ratio; it's how many steps you run per month.

To calculate your own numbers, run /usage in Claude Code to get four token counts, then paste them into this:

# 用 token 數回推一次 Claude Code 工作階段的實際花費。
# 在 Claude Code 裡執行 /usage 就能拿到這四個真實數字。
RATES = {  # Kunavo 每 1M token 的美元價：(input, output)
    "claude-haiku-4-5":  (0.70, 3.50),
    "claude-sonnet-4-6": (2.10, 10.50),
    "claude-sonnet-5":   (1.40, 7.00),
    "claude-opus-5":     (3.50, 17.50),
}

def session_cost(model, inp, out, cache_read=0, cache_write=0):
    i, o = RATES[model]
    return (
        inp / 1e6 * i                   # 新的輸入
        + out / 1e6 * o                 # 輸出，含思考 token
        + cache_read / 1e6 * i * 0.10   # 讀快取：輸入費率的 10%
        + cache_write / 1e6 * i * 1.25  # 寫 5 分鐘快取：125%
    )

print(session_cost("claude-sonnet-5", 1_200, 5_300, 940_000, 50_000))

Paying from Taiwan

Both routes in Taiwan use an international card; the difference is how many times it is charged. The API route is prepaid: top up with an international credit card (Visa, Mastercard, Amex, JCB, UnionPay), Apple Pay, or Google Pay, with a minimum of $10 and no balance expiration, then deductions are made according to usage — the fee occurs only at the time of top-up. claude.ai subscriptions are recurring charges in U.S. dollars, and the card issuer's international transaction fee is charged again every month.

Taiwan currently has no local payment options—JKO Pay and LINE Pay aren't on the available list, so there's no need to look for a local payment method from Taiwan: there isn't one. Local payment methods vary by country (Brazil's Pix, Belgium's Bancontact, Poland's BLIK, and others), and Taiwan isn't included. See Billing documentation for the full list.

Connect Claude Code to a cheaper endpoint

Claude Code uses two environment variables to decide where to send requests. Pointing it to another endpoint requires no change to the installation or CLI:

export ANTHROPIC_BASE_URL=https://api.kunavo.com
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

The model lines cannot be omitted: Claude Code's default model and the opus alias both point to the latest Opus, so if Kunavo has not listed it yet, the first request will receive a 404. Here, pin the opus alias to Opus 5.5 (claude-opus-5-5); Claude Code v2.1.280 or later is required, and older versions should first run claude update. The sonnet alias points to Sonnet 5.5, which Kunavo does not provide — without setting ANTHROPIC_DEFAULT_SONNET_MODEL, the execution stages of /model sonnet and opusplan, as well as subagents set to model: sonnet, will receive a 404. After the change, Claude Code continues to work normally, but model calls use the new endpoint and rates. To return to the subscription, remove these variables. See Claude Code API key setup for configuration details; to route different requests to different models, see Claude Code Router.

Four ways to lower Claude Code costs

1. Make caching work.Cached input costs just 10% of the rate, and Claude Code rereads the entire conversation at every step (how it works).
2. Clear context between unrelated tasks.A session that never ends resends its entire history at every step; shorter context costs less.
3. Choose a model to fit the problem.Making the largest model your default is the most expensive and most common mistake.
4. Do not remove the Haiku line. Without ANTHROPIC_DEFAULT_HAIKU_MODEL, the haiku alias requests a model name that Kunavo does not offer (Haiku 5.5 from v2.1.293 onward), so /model haiku and subagents set to haiku both return 404s. The same line also moves the many small internal calls from the main model to Haiku.

If you haven't installed it yet, see Claude Code installation guide. For daily use and switching models after installation, see Claude Code tutorial. To use Codex on the OpenAI side instead, see Codex tutorial.

Frequently asked questions