Claude Prices 2026 — What Does Claude Cost? Subscription, API, and Claude Code

How much does Claude cost? There are three separate prices, depending on how you use Claude. Chat at claude.ai is free on the Free plan, Claude Pro costs $20/month, and Claude Max starts at $100/month. If you integrate Claude into your own software through the API, there is no subscription; you pay per token—on Kunavo, starting at $0,70 per 1M input tokens, which is 30% below Anthropic’s list price. And Claude Code has no separate price: it runs either on a subscription or with an API key. This page breaks down all three levels.

Prices checked on October 3, 2026 against claude.com/pricing and Anthropic’s API pricing documentation. This page reads Kunavo rates live from the model catalog—what you see here is what the API charges.

Claude’s three pricing levels at a glance

The two options cannot be combined: as long as an access variable such as ANTHROPIC_AUTH_TOKEN is set, Claude Code uses the key, the subscription remains idle, and usage is charged to the key owner.

1. Claude subscription prices—Free, Pro, Max, and Team

This is the price most people mean when they search for Claude AI cost: the plan fee for chat at claude.ai. Anthropic publishes it at claude.com/pricing, and bills in US dollars.

Claude Pro cost—and when Max is worth it

Claude Pro costs $20 per month, or $17 per month with annual billing ($200 upfront). The more interesting part of subscriptions is not the fee but how usage works: it is measured against a rolling five-hour window and a weekly window, and both windows are shared between chat and Claude Code. An afternoon in the web app therefore draws from the same allowance as the coding session in the evening. Switching models does not reset either window, because the windows apply across all models.

That is why the honest answer to “is Pro enough for me?” is not a pricing question but a question about how you work. Anyone who spends a few hours a day working with Claude will generally stay within Pro. You know it is time to move to Max, starting at $100/month, when you repeatedly hit the limits—not because you calculated it in advance.

Is Claude free?

Claude chat is free on the Free plan; the Claude API is not. According to Anthropic, the Free plan provides about half the weekly allowance of a paid plan and does not include Claude Code. There is no permanent free tier for the API: Anthropic’s pricing FAQ says new users get a small amount of starter credit for testing, after which billing is per token. So the least-committing option is not a free allowance but pay-as-you-go—on Kunavo, with a $10 top-up and no monthly fee—and a quiet month really costs nothing because the balance does not expire.

2. Claude API prices—rates by model

The API has no plans or usage windows: you pay per token, with input and output billed separately. Rates are in USD per million tokens, as charged by Kunavo; the “Anthropic list” column shows the published price for the same model.

Live rates are always available on the pricing page and on each model page (claude-sonnet-4-6, claude-haiku-4-5, claude-opus-5, claude-opus-5-5). Haiku is the budget option, Sonnet 4.6 is the previous generation, Sonnet 5 is the current Sonnet (Kunavo’s least expensive), Opus 5.5 is the newest Opus tier—cheaper on Kunavo than Opus 5—and Fable 5 is Anthropic’s most capable model.

Claude Opus 5.5 is now the best deal among Opus models

Anthropic released Claude Opus 5.5 on September 22, 2026, listing it at $4,00 / $20,00 per 1M tokens—below Opus 5 at $5,00 / $25,00. According to Anthropic, it performs at the level of Claude Fable 5.1 for most tasks. On Kunavo, Claude Opus 5.5 costs $2,80 / $14,00 per 1M tokens, around 30% below Anthropic’s rate—Kunavo’s lowest Opus rate and less than Opus 5 ($3,50 / $17,50). Sonnet 5 is now even cheaper at $1,40 / $7,00. A long-context analysis with 50,000 input and 2,000 output tokens costs $0,42 on Fable 5, $0,21 on Opus 5, and $0,17 on Opus 5.5. Adaptive thinking is always active on Opus 5.5, so include thinking tokens in the output line.

Claude Opus 5 remains available. Anthropic lists it at $5,00 / $25,00—half the price of Fable 5 ($10,00 / $50,00) with nearly equivalent reasoning. On Kunavo, it costs $3,50 / $17,50 per 1M tokens, around 30% below Anthropic’s rate and exactly the rate we charge for the older claude-opus-4-7: moving a workload from Opus 4.7 to Opus 5 means changing one word in the model name, with no extra cost.

Claude Sonnet 5—now also below Anthropic’s rate

Anthropic had announced $2,00 / $10,00 per 1M tokens as the introductory price for Claude Sonnet 5, ending on August 31, 2026—and has since made this the permanent standard price; the increase to $3.00 / $15.00 planned for September 1, 2026 will not happen. Kunavo ($1,40 / $7,00): around 30% below Anthropic's rate.

That leaves the practical question: Sonnet 5 or Opus? claude-opus-5-5 costs $2,80 / $14,00 on Kunavo—the cheapest Opus here, with Opus reasoning. Sonnet 5 itself is even cheaper on Kunavo at $1,40 / $7,00; if you are weighing the two, these are the two rows that matter.

Cost examples you can calculate

Real figures calculated using Kunavo rates:

10,000 support tickets per day on Haiku 4.5 cost around $12,60 per day. Here is the calculation so you can run it yourself:

# Claude-Tarife auf Kunavo (USD pro 1M Tokens):
# (Input, Output, Cache-Treffer als Anteil des Input-Tarifs)
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),
    "claude-fable-5":    (7.00, 35.00, 0.10),
}

def kosten(model: str, in_tokens: int, out_tokens: int, cache_read: int = 0) -> float:
    i, o, c = RATES[model]
    return (
        in_tokens / 1_000_000 * i          # frischer Input
        + out_tokens / 1_000_000 * o       # Output, inkl. Thinking-Tokens
        + cache_read / 1_000_000 * i * c   # Cache-Treffer: Anteil c des Input-Tarifs
    )

print(kosten("claude-haiku-4-5", 800, 200))      # ein Support-Ticket   -> $0.00126
print(kosten("claude-sonnet-4-6", 6_000, 500))   # eine RAG-Antwort     -> $0.0179
print(kosten("claude-opus-5", 50_000, 2_000))    # Long-Context-Analyse -> $0.21
print(kosten("claude-opus-5-5", 50_000, 2_000))  # Analyse auf Opus 5.5 -> $0.168

Two surcharges that affect German budgets

Both are often overlooked, and both push costs in the same direction. First: German uses around 20–30% more tokens than English for the same content—compound words, longer words, and finer tokenizer segmentation. Second: Anthropic notes that Claude 4.7 and newer models (including Opus 5 and Fable 5) use a new tokenizer that produces around 30% more tokens for the same text than the tokenizer in Sonnet 4.6 and earlier. So if you are planning a German-language product on one of the newer models, do not rely on token counts from an English blog post using Sonnet 4.6; measure it yourself once.

Prompt caching—the biggest cost lever

If your system prompt or retrieved context is large and stable, prompt caching charges cached input at 10% of the input rate (2,5% bei Claude Fable 5.1, 5% bei Claude Opus 5.5). For a long, repeated system prompt, this cuts input costs by 60–90%. Writing to the cache costs more upfront—1.25x the input rate with a five-minute TTL, or 2x with a one-hour TTL—and pays for itself after the first or second cache hit, respectively. It is one field in the request; details are in the caching documentation and the Messages API reference.

3. Claude Code costs—subscription or API key

Claude Code itself is free: the CLI, VS Code and JetBrains extensions, hooks, skills, subagents, and MCP servers are free. You pay only for model calls—and that is where the Claude Code price is determined: either through the subscription (included in Pro and Max, subject to usage windows) or through an API key (per token, with no monthly fee or usage window).

What an agentic step costs

Costs depend on the number of agentic steps, not calendar days. A step is one model round trip within a task: system prompt, conversation history, and fresh file context go in; an edit or explanation comes out—roughly 25,000 input and 1,200 output tokens. Your settings determine which model handles the step; unless configured otherwise, Claude Code uses its default model, Opus 5.5. At Kunavo rates, without caching:

These figures are deliberately pessimistic: no caching and one model for everything. In a warm session, most input tokens are cache hits at 10% of the input rate (bei Opus 5.5 zu 5%), which is why a naive “tokens × price” calculation significantly overestimates the bill. For context from actual installations, Anthropic’s own cost documentation cites around $13 per developer on an active day and $150–250 per developer per month, with 90% of users staying under $30 per active day—at list prices and with mixed model costs.

The break-even point between a subscription and a key

Putting both options on the same scale, at $0,0434 per step on Sonnet 5:

• $20 (Pro) covers around 460 agentic steps per month—about 23 tasks of 20 steps each.
• $100 (Max, entry tier) covers around 2.300 steps—about 115 tasks.

The rule is simple: if you complete significantly more than two dozen real tasks every month, the flat fee is a good fit. If your work comes in bursts—an intensive week followed by two quiet ones—pay-as-you-go is better, because quiet weeks cost nothing. Teams almost always want keys anyway: they are the only way to allocate usage per person and enforce hard limits.

Switch Claude Code to Kunavo

Claude Code reads ANTHROPIC_BASE_URL natively and therefore works with any endpoint that serves the Anthropic Messages API—no plugin, proxy, or patched binary required:

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

ANTHROPIC_BASE_URL is only the origin; Claude Code appends the /v1/messages path itself. The model lines are part of this: Claude Code's default model and the opus and sonnet aliases follow Anthropic's latest models, and if Kunavo does not yet offer one of them, the very first request ends with a 404. The sonnet alias requests Sonnet 5.5, which Kunavo does not offer — without ANTHROPIC_DEFAULT_SONNET_MODEL, /model sonnet, the execution phase of opusplan, and every subagent using model: sonnet end with a 404. The opus alias sets the block to Opus 5.5 (claude-opus-5-5); this requires Claude Code v2.1.280 or later, while claude update updates older installations. Use ANTHROPIC_AUTH_TOKEN instead of ANTHROPIC_API_KEY: the two go in different headers, and the latter requires one-time interactive approval that never appears again after rejection. Create the key (sk-kn-…) after signing up in the Dashboard. The detailed version is in the Claude Code Router Guide.

If you use Claude through a coding agent such as Kilo Code, Cline, or Roo Code, configure the same settings everywhere: provider “OpenAI Compatible,” base URL https://api.kunavo.com/v1, your key, and a Claude model slug—the step-by-step guide is Set up Kilo Code with the Claude API.

Which Claude tier should you choose? Comparing costs

The question of which tier to choose almost always comes with a pricing question, and the answer depends on what you need most.

At the lower end, Claude Haiku 4.5 costs $0,70 per 1M input tokens, around 67% below Claude Sonnet 4.6. The straightforward rule is: Haiku for high volume and simple tasks, Sonnet for code, reasoning, and agents, and Opus for the difficult minority. The difficulty-based routing pattern (simple tasks go to the cheaper model) is covered in the AI cost optimization guide. All models are available through one Kunavo key and the same balance—including GPT, image, and video models—with no provider invoices to reconcile.

Pay for Claude from Germany

Payment is processed through Stripe. There is no subscription or Anthropic Console setup: you top up a balance, and each call is charged at the per-token rates above. Available in Germany:

• Cards: Visa, Mastercard, American Express (credit and debit cards)
• Apple Pay, Google Pay, Link (one-click payment)

SEPA Direct Debit does not appear in the checkout. A debit card linked to a current account works normally. Pay by Bank is available in the checkout for buyers in the United Kingdom, Ireland, and Finland, but not Germany: there, it is still a private preview in Stripe (Stripe documentation on Pay by Bank, read on October 3, 2026). Stripe may convert the dollar amount to euros in the checkout; if you pay in USD, your bank performs the conversion.

$10 minimum top-up, pay-as-you-go, and since it is a prepaid balance, the balance never expires. Regarding VAT: the standard checkout does not show VAT and has no VAT ID field; an invoice showing VAT is not issued automatically. Your top-ups appear in the billing dashboard with their date, amount, and status; if you need a receipt for your records, email contact@kunavo.com. For German-specific details—latency from Frankfurt, GDPR, and the legal notice—see the Claude API guide for Germany.

Reduce Claude costs—in this order

1. Start with prompt caching. Cached input costs 10% of the input rate (2,5% bei Claude Fable 5.1, 5% bei Claude Opus 5.5)—the best return for the effort, and by far the biggest impact in Claude Code because every step resends the full conversation history.
2. Clear context between unrelated tasks. Context is sent with every step, so an all-day session gets quadratically expensive. A fresh task starts with a small context again.
3. Choose a model for the task. Leaving the largest model as the default is the most common cause of an unexpectedly high bill. Use Haiku for simple work and Opus for the difficult minority.
4. Cap output. Set max_tokens—output costs about five times as much as input.
5. Retrieve less input. Use a few precise RAG chunks instead of half the knowledge base.

Together, these levers typically cut a Claude bill in half or more. For a gateway comparison, see Kunavo vs. OpenRouter.

Frequently asked questions