Claude pricing 2026 — subscription prices, API rates, and the break-even point

Claude costs fall into three categories.Chat subscriptions on claude.ai: Free is $0, Claude Pro is $20/month ($17/month when billed annually), and Claude Max starts at $100/month. The API has no monthly fee and is billed per token. On Kunavo, Claude Sonnet 4.6 costs $2.10 / $10.50 per 1M input / output tokens, compared with Anthropic’s official prices of $3.00 / $15.00 — 30% lower, depending on the model. Claude Code has no separate price; it runs on one of the two above.

Confusion over “how much Claude costs” almost always comes from mixing up these three categories. This page lists current prices for each one, then works out which is cheaper for the same usage: a subscription or the API. Lots of pages list the plans; very few calculate the break-even point.

Rates verified on October 3, 2026: subscription prices are from claude.com/pricing. Kunavo’s figures come from this site’s catalog and match the amounts actually billed.

Three types of costs, in one table

1. Claude subscription prices — Free / Pro / Max / Team

claude.ai subscriptions bundle chat, Claude Code, and various features into a monthly fee with included usage limits. Limits are managed through a 5-hour rolling window and weekly caps, and chat and Claude Code share the same window — a chat session in the afternoon and coding in the evening both use the same allowance.

The decision to upgrade to Max depends on how often you hit the limits.If you are blocked several times a week and have to stop working, the $80 difference pays for itself in time saved. If it happens only once a month, using an API key to finish that task will cost less. For options after hitting a limit, see Claude Code costs; for how the limits work and when they reset, see Claude Pro usage limits.

2. Claude API costs — per 1M tokens by model

The API is billed per token, with separate rates for input and output; output costs more. Below are the rates per 1M tokens as of October 3, 2026.

Claude Opus 5.5 pricing — the new Opus that costs less than Opus 5

Claude Opus 5.5 is Anthropic’s latest Opus, released on September 22, 2026. Its official price per 1M tokens is $4.00 for input and $20.00 for output, lower than Opus 5 at $5.00 / $25.00. Anthropic says it performs at the level of Claude Fable 5.1 on most tasks. On Kunavo, it costs $2.80 / $14.00, about 30% below the official price. It costs more than Kunavo’s Sonnet 5 ($1.40 / $7.00), but less than Opus 5 ($3.50 / $17.50) — it has the lowest rate in this table for Opus-level reasoning.

Context window of 1M tokens, with up to 128K output tokens per request. Adaptive thinking is always on, and thinking tokens are billed at the output rate, so include them in the output estimate. Cached reads cost only 5% of the input rate, lower than the 10% for the other models in this table. The model ID is claude-opus-5-5; see the Claude Opus 5.5 model page for rate details.

Claude Sonnet 5 pricing — also below the official price since September 24

Since September 24, 2026, Kunavo’s price for Claude Sonnet 5 has also been below the official price.After Anthropic made the $2.00 / $10.00 rates originally announced as introductory prices its standard rates, Kunavo’s price, which had matched them, dropped to $1.40 / $7.00, about 30% below the official price. For Opus-level reasoning, Claude Opus 5.5 ($2.80 / $14.00) is still the least expensive in this table. For tasks that only need Sonnet-level performance, Sonnet 5 costs less.

Calculate it yourself

It’s more useful to plug your own usage into the rate table. If you don’t want to write code, the Claude token calculator lets you enter input, output, and cached tokens to calculate per-call and monthly costs. To work it out yourself, calculate one month directly here:

# 一個月的 Claude API 費用，直接用實際 token 數算。
# 每 1M token USD：(input, output, 快取讀取佔輸入費率的比例)
RATES = {
    "claude-haiku-4-5":  (0.70, 3.50, 0.10),
    "claude-sonnet-4-6": (2.10, 10.50, 0.10),
    "claude-opus-5":     (3.50, 17.50, 0.10),
    "claude-opus-5-5":   (2.80, 14.00, 0.05),
}

def monthly_cost(model, inp, out, cache_read=0, cache_write=0):
    i, o, r = RATES[model]
    return (
        inp / 1e6 * i                   # 新輸入
        + out / 1e6 * o                 # 輸出（含 thinking token）
        + cache_read / 1e6 * i * r      # 快取讀取：輸入費率的 10%，Opus 5.5 是 5%
        + cache_write / 1e6 * i * 1.25  # 5 分鐘快取寫入：125%
    )

print(monthly_cost("claude-sonnet-4-6", 6_000_000, 2_000_000, cache_read=24_000_000))

Caching has the biggest impact

Cached reads cost only 10% of the input rate (Claude Opus 5.5 is 5%).For long system prompts, documents read repeatedly, and conversation history, whenever the same content is sent repeatedly, cached reads alone provide the largest cost reduction. Writes cost 125% of the input rate for a 5-minute cache, so the content needs to be read a second time to be worthwhile. Enabling caching for a prompt whose content changes every time will therefore cost more. For details on how it works and common reasons caching stops working through a gateway, see Claude prompt caching and Caching documentation.

3. Claude Code costs — subscription or API key

Claude Code itself has no price; the CLI, editor extension, and settings are free. You pay only for model calls. The deciding factor is how many times you run it, not how many days you work: one agentic step on Claude Sonnet 4.6 costs about $0.0651, so $20/month covers about 310 steps, or 15 tasks with 20 steps each. For chat, $20/month covers about 1,700 turns. When run with an API key, Claude Code currently defaults to Claude Opus 5.5, for which a step of the same size costs about $0.0868.

For the full calculation, environment variable setup, and how to continue after reaching the usage limits, see Claude Code costs. To compare side by side with Codex using the same unit, see Codex costs.

Paying from Taiwan

Subscriptions are recurring charges denominated in US dollars and require a credit card that can make international transactions. Kunavo API balances use Stripe and support cards including JCB (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, Google Pay, and Link. It is prepaid, with a minimum top-up of $10; balances do not expire, and failed requests are not charged.

Taiwan has no local payment methods — JKoPay and LINE Pay are not on the list of available methods. Both options require an international card. The difference is frequency: a subscription is billed monthly and incurs an overseas transaction fee each time; prepaid balance is charged only when you top up. For billing details, see the Billing documentation; rates for all models are listed on the model list.

How to reduce Claude costs — ordered by impact

1. Enable prompt caching. Cached reads cost 10% of the input rate (5% for Opus 5.5). This has the biggest impact when you repeatedly send the same content.
2. Choose a model based on the task. Don’t set the most expensive model as the default. Haiku is usually enough for classification, extraction, and formatting. The difference between $0.0217 and $0.11 per step goes straight to your bill.
3. Limit output. Output costs more than input, so max_tokens and specifying an output format are effective.
4. Don’t add unrelated history. Start a new conversation for a new task — long histories are billed as full input every time.

After these four steps, switching to an endpoint priced below the official rates compounds the savings. Kunavo provides both OpenAI-compatible and Anthropic-compatible APIs, so existing code can work by changing only base_url. See the Quickstart for instructions.

If your card is declined or you don’t have a credit card, see Claude API payment methods.

Frequently asked questions