Claude Code pricing and costs — subscription or API: at what usage each option pays off

Claude Code itself costs nothing — you pay for the model behind it, and there are two ways to do that. With a subscription, Claude Code is included in the monthly fee: Claude Pro costs $20 per month ($17 per month with annual billing, $200 upfront), while Claude Max costs $100 or $200. These are Anthropic’s US prices before tax; prices and currencies vary by region. Private customers in Germany pay 19 percent value-added tax, calculated by Anthropic based on the billing address. The free plan doesn’t include Claude Code. With an API key, there’s no base fee, just a per-token price: on Kunavo, Claude Sonnet 5 costs $1,40 per million input tokens and $7,00 per million output tokens, so a typical agent step costs around $0,0434 (about 0,038 €). Anthropic itself estimates enterprise usage at about $13 per developer per active day.

Choosing the cheaper option comes down to division: the $20 Pro fee is equivalent to about 460 such steps per month on Kunavo. If you use Claude Code almost every day and stay within Pro’s limits, Pro is usually the better deal. If your usage comes in bursts, the key costs less — a quiet month then costs $0.

As of September 23, 2026. Anthropic’s figures come from claude.com/de/pricing, Anthropic’s Help Center, and the API pricing documentation; Anthropic may change its prices and plans. Kunavo rates come directly from Kunavo’s model catalog. Euro amounts are converted using the ECB reference rate from September 23, 2026 (1 € = $1,1411) and are indicative.

How much does Claude Code cost? A guide to subscriptions and API pricing

The two do not offset each other. While an access variable such as ANTHROPIC_AUTH_TOKEN is set, Claude Code uses the key; your subscription login remains stored but unused, and usage is billed to the key. Remove the variable and the subscription is active again. /status shows which route is currently in use.

Claude Code with a subscription: Pro, Max, and Team

According to Anthropic's setup documentation, Claude Code works with a Pro, Max, Team, or Enterprise subscription, a Console account for the API, or a cloud provider such as Amazon Bedrock; the free plan does not include it. There is also the LLM gateway route, which this article demonstrates with Kunavo below. Prices from claude.com/de/pricing and the Help Center, in US dollars before tax, as of September 2026:

Anthropic says Max prices apply to web subscriptions; prices in mobile apps may differ. According to the Help Center, Fable models are included in the allowance on Max and Team Premium seats, up to 50 percent of the weekly limit; on Pro and Team Standard seats, they use usage credits at API prices only, including in Claude Code.

With or without VAT?

Anthropic's prices in dollars are before tax. The German pricing page says beneath the plans: “Prices shown do not include any applicable taxes.” Anthropic calculates tax based on the billing address for the payment method. For private customers in Germany, VAT is 19 percent (Section 12 of the German VAT Act), because electronic services to private customers are taxed at the customer's place of residence (Section 3a(5) of the German VAT Act).

Checkout is the only way to see what you will actually pay in Germany. According to the Help Center, the Pro plan costs $20 per month in the US; monthly prices vary by region, are shown in local currency where supported, and in some regions tax is already included in the displayed price, while in others it is added at checkout. The Help Center does not state a price in euros; Anthropic shows current prices for your region at claude.ai/upgrade. If you subscribe on the web, you receive an invoice by email after each charge and, depending on your location, can add a VAT ID that applies to future invoices. If you subscribe in the iOS or Android app, Apple or Google handles billing.

The limits are the real price

Pro and Max measure usage with a session limit that resets every five hours and a weekly limit across all models. Both are shared between Claude chat and Claude Code: everything you do in both tools counts against the same allowance. Anthropic does not publish fixed figures for these limits. For a detailed explanation of how the windows work and what to do when you hit a limit, see Claude Limits.

If you want to keep working after hitting a limit, you can enable usage credits. According to Anthropic, these are prepaid and billed at standard API prices. So beyond your limit, the subscription already charges you the API price—and by default, the cache then lasts only five minutes instead of 60. This is where your own key comes in: Claude models on Kunavo cost 30% less than Anthropic's, depending on the model.

Is Claude Code free?

The tool is free; usage is not. The CLI and extensions for VS Code and JetBrains are free, and there is no extra charge for hooks, skills, subagents, or MCP. You pay for model calls—including those triggered by a subagent—and there is no free way to make them: the Free plan does not include Claude Code, and the API charges from the first token. According to Anthropic, new users of the Anthropic API receive a small amount of starter credit for testing, but there is no ongoing free tier.

The least-commitment option is a prepaid key: on Kunavo, you can top up from $10, with no monthly fee, and the balance does not expire. With Claude Sonnet 5, $10 covers around 230 agentic steps, or about 12 tasks of 20 steps each, without cache.

Claude Code through the API: per-model pricing

With a key, there is no plan or usage window: Claude Code is billed per token, like any other API traffic. Files and tool outputs count as input; edits, explanations, and thinking tokens count as output. Prices in US dollars per million tokens (input / output):

claude-opus-5 costs $3,50 / $17,50 on Kunavo.

One detail can skew price comparisons: according to Anthropic, Claude models from version 4.7 onward, including Claude Sonnet 5, Claude Opus 5, and Claude Fable 5, use a new tokenizer that produces around 30 percent more tokens for the same text than Claude Sonnet 4.6; the exact increase depends on the content. So the same work costs a little more on the newer models than the rates alone might suggest.

Caching determines the cost

Claude Code sends the entire conversation with every step. Most of the input is therefore repeated context—and prompt caching is what makes that cheaper. Here is how Kunavo bills Claude Sonnet 5:

The coffee break costs money. On an API key, the cache lasts five minutes by default, according to Anthropic's cost documentation; on a subscription, it lasts one hour. If you take a longer break, the next step charges you to send the entire context again, as a cache write rather than a cache hit. With 200,000 tokens of context on Claude Sonnet 5, that reread alone costs around $0,35 on Kunavo instead of $0,028. So if you are starting something new after a long break anyway, it is better to begin with /clear.

How much does Claude Code cost per month through the API?

Costs follow agentic steps, not workdays. A step is one back-and-forth with the model within a task: system prompt, conversation history, and fresh file context go in, and an edit or explanation comes out—roughly 25,000 input tokens and 1,200 output tokens. At Kunavo prices, calculated deliberately without cache:

Twenty such days on Claude Sonnet 5 cost around $87 (about 76 €); at Anthropic's list price, it would be $124. With a warm cache—assuming 20,000 of the 25,000 input tokens come from cache and the rest are written anew—a step drops to $0,0200 and the month to around $40.

For a real-world comparison, Anthropic's cost documentation gives averages for enterprise deployments of around $13 per developer per active day (about 11,39 €) and $150–250 per developer per month (about 131–219 €); nine out of ten users stay below $30 per active day. These are averages from enterprise deployments, not Kunavo prices; Anthropic recommends measuring your needs first with a small pilot group. You can calculate your own scenarios with the Claude Token Calculator.

Subscription or API: the break-even point in euros

Both options on the same scale: monthly fee ÷ cost per step = number of monthly steps at which the subscription pays off. Calculated using Anthropic's dollar prices before tax and without cache:

In euros: One step on Claude Sonnet 5 costs about 0,038 € through Kunavo. That means every euro of subscription fees covers around 26 steps. Multiply the euro price Anthropic shows you at checkout by 26—that is your break-even point in steps per month.

With a warm cache, the break-even point moves higher: $20 buys around 1.000 steps on Claude Sonnet 5. Anthropic does not publish how many steps each subscription plan includes; the table only shows when the flat fee becomes cheaper—provided the limits allow that much usage.

The takeaway: Pro pays off quickly. Around 5 busy days (460 steps) on a key already cost as much as $20—if you use Claude Code almost every day, Pro is cheaper, as long as its limits are enough. For Max, the break-even point is much higher: Max 5x on Claude Sonnet 5 only pays off at around 2.300 steps per month, or around 23 busy days; on Claude Sonnet 4.6, it takes around 1.540 steps. If you work in bursts, you stay below these break-even points and pay less with a key, because quiet weeks cost nothing. Teams can use one key per person or CI job, each with its own spending limit.

Is Claude Max worth it?

Max 5x costs $80 more than Pro. For the same amount, a key on Kunavo gets you around 1.840 steps with Claude Sonnet 5 (on Claude Sonnet 4.6, around 1.230), without cache—that is about 37 afternoons blocked by limits, at 50 steps per afternoon, where each step costs around $2,17. Purely by the numbers, Max 5x only pays off if the work Pro blocks you from doing exceeds that amount. If you hit the limit less often, it is cheaper to keep Pro and bridge those hours with a key. Claude Limits explains how to keep your subscription and use a key only for the hours when you are blocked.

Recalculate a session using your own figures

/usage shows the token counts for the current session in Claude Code, broken down into fresh input, output, cache reads, and cache writes. Claude Code calculates the dollar figure beside them locally at list price; with Kunavo, the actual charge is lower for discounted models. Use the four figures to recalculate it:

# Was eine Claude-Code-Sitzung gekostet hat – aus ihren eigenen Token-Zahlen.
# /usage in Claude Code zeigt die vier Werte der laufenden Sitzung.
PREISE = {  # USD pro 1M Tokens auf Kunavo: (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),
}
CACHE_LESEN, CACHE_SCHREIBEN = 0.10, 1.25  # Faktoren auf den Input-Preis

def sitzungskosten(modell, neu, output, cache_lesen=0, cache_schreiben=0):
    i, o = PREISE[modell]
    return (
        neu * i                                   # frischer Input
        + output * o                              # Output inkl. Thinking-Tokens
        + cache_lesen * i * CACHE_LESEN           # Cache-Treffer
        + cache_schreiben * i * CACHE_SCHREIBEN   # Cache-Schreibvorgänge
    ) / 1_000_000

# Anthropics Beispielsitzung aus der /usage-Dokumentation:
print(round(sitzungskosten("claude-sonnet-4-6", 1_200, 5_300, 940_000, 50_000), 2))  # 0.39

Anthropic's example session from the documentation runs on Claude Sonnet 4.6: 1,200 fresh input tokens, 5,300 output, 940,000 from cache, and 50,000 written. It costs $0.55 at list price and $0,3868 through Kunavo. The 940,000 cache tokens alone cost $0,20 instead of $1,97 without cache: Cache savings in this session exceed the cost of the entire session.

Five ways to reduce costs

1. Let the cache do its job. A cache hit costs only a fraction of the input price for the models on this page, and Claude Code rereads the entire conversation at every step (how caching works).
2. Clear context between tasks. Context is sent with every step, so an all-day session gets more expensive with every message. /clear and starting a new task resets the context to a smaller size.
3. Choose the model for the task. According to Anthropic, Sonnet is enough for most coding tasks; Opus is for architectural decisions and multistep reasoning. A step on Claude Fable 5 costs around 5,0 times as much as one on Claude Sonnet 5.
4. Use Haiku for small tasks. According to Anthropic, ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5 determines which model the haiku alias and background tasks such as session titles use—here, Claude Haiku 4.5, the cheapest model on this page. You can also route simple subagents there by using model: haiku in their configuration.
5. Adjust thinking effort. Thinking tokens are billed as output, and the default budget can be tens of thousands of tokens per request, depending on the model. For simple tasks, /effort reduces the effort.

Set up Claude Code with Kunavo

Claude Code reads ANTHROPIC_BASE_URL natively and works with any endpoint that serves the Anthropic Messages API—no plugin, proxy, or reinstall required. Set it in the shell, once in the terminal, or persist it in ~/.zshrc:

export ANTHROPIC_BASE_URL=https://api.kunavo.com   # nur der Origin, ohne /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

ANTHROPIC_BASE_URL is only the origin; Claude Code appends the /v1/messages path itself, and a value ending in /v1 leads to a 404. Use ANTHROPIC_AUTH_TOKEN instead of ANTHROPIC_API_KEY: Kunavo accepts the key in both headers, but ANTHROPIC_AUTH_TOKEN takes effect immediately, while Claude Code initially asks for confirmation of ANTHROPIC_API_KEY in interactive mode.

The model lines are also part of this. Without ANTHROPIC_MODEL, Claude Code starts on its default model, currently Opus 5.5 according to Anthropic; this line instead makes Claude Sonnet 5 the starting model used for all calculations on this page. ANTHROPIC_DEFAULT_OPUS_MODEL sets the alias opus to Claude Opus 5.5 (claude-opus-5-5, on Kunavo $2,80 / $14,00 per million input and output tokens). Claude Code requires at least version 2.1.280 for Opus 5.5; claude update brings an older installation up to date. The sonnet alias instead requires Sonnet 5.5, which Kunavo does not offer: Without ANTHROPIC_DEFAULT_SONNET_MODEL, /model sonnet, the execution phase of opusplan, and subagents with model: sonnet end with a 404. The line redirects the alias to Claude Sonnet 5. Permanently, the variables belong in the env block of your user settings:

{
  "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"
  }
}

Never put them in the project's .claude/settings.json—that ends up in the repository. For the VS Code extension, put the variables in claudeCode.environmentVariables in your VS Code user settings. To check, /status shows an “Anthropic base URL” line with the Kunavo endpoint and a line for the auth token; if it shows a claude.ai login instead, the variable was not loaded. The full guide (in English): Claude Code with Kunavo.

Create the key (sk-kn-…) in the Dashboard after signing up. Payment is made through Stripe as prepaid credit starting at $10: by card (Visa, Mastercard, American Express) or Link; in Safari and on iOS, also with Apple Pay; in Chrome and on Android, with Google Pay (in each case using a card stored in the wallet); buyers in Austria also see EPS. SEPA direct debit and PayPal are not available. Pay by Bank appears in Checkout only for buyers in the United Kingdom, Ireland, and Finland, so not in Germany or Austria; for Germany, Stripe still offers it as a private preview (as of: October 3, 2026). Visa or Mastercard debit cards work like credit cards. Prices are in dollars, and Checkout displays the amount in your currency. The credit does not expire, failed requests cost nothing, and each key can have a monthly spending limit: once it is reached, the API rejects further requests with 402. All top-ups and deductions appear in the history under /app/billing.

What changes with a key—and what Kunavo does not replace

The core of Claude Code—the agent, tools, subagents, and MCP servers—runs through the Messages API and stays the same. Remote Control and voice dictation, however, require a claude.ai identity and are not available with a gateway key.

Kunavo serves Claude through shared capacity, without dedicated capacity or a contractual SLA. If you need guaranteed capacity, contract directly with Anthropic. And if you code every day and can work within Pro's limits, Pro is probably cheaper. A key pays off for bursty use, teams with per-person budgets, CI, and hours beyond the limit. For the total cost of Claude—chat, API, and Claude Code side by side—see Claude pricing; for the same calculation for OpenAI's agents, see Codex costs.

Frequently asked questions