Claude Code pricing 2026—subscription vs. API, per-model rates, and payment with Pix

Claude Code has two completely separate prices, and nearly all confusion about its cost comes from mixing them up. A claude.ai subscription includes the tool in a fixed monthly fee with usage limits. An API key charges per token, with no monthly fee or usage limit. This page gives the current figures for both options, the calculation to help you choose, the cost of a real month of use, and — the part often missing from English-language coverage — how to pay from Brazil.

Rates verified on October 3, 2026 against claude.com/pricing and Anthropic’s API pricing documentation. Kunavo’s per-token prices are read live from the model catalog.

The two options, side by side

Subscription pricing

Claude Code is included in claude.ai paid plans — Anthropic’s installation documentation says the Free plan doesn’t include access to Claude Code. Pro costs $20/month ($17/month with annual billing, $200 upfront) and suits one developer working within the plan limits. Max starts at $100/month and is for people who keep hitting those limits. Team and Enterprise add per-seat allowances and administrative spending controls.

Two details matter in Brazil: billing is in dollars and recurring, on an international credit card enabled for foreign transactions — with IOF and the issuer’s international transaction fee charged every month — and usage limits are shared with Claude chat, so an afternoon of conversation uses the same window as your coding session.

API pricing

There’s no monthly fee here: you pay per token. Rates in USD per 1M tokens, with Anthropic’s official table alongside for direct comparison.

“Cost per step” means one round trip to the model within a task — system prompt, history, and file context in the input (~25,000 tokens), with an edit or explanation in the output (~1,200). It’s the right unit for thinking about Claude Code because the tool rereads the context at every step.

How much a real month costs

Anthropic’s cost documentation reports about $13 per developer per active day and $150–250 per developer per month for enterprise deployments, with 90% of users below $30 per active day. These are figures at full list price using a mix of models. At Kunavo rates, the reduction in cost for the same volume is 30% — and routing background calls to Claude Haiku 4.5 ($0.70 / $3.50 per 1M) cuts it further.

At Kunavo rates, the break-even point is: $20 buys approximately 460 agentic steps per month with Sonnet 5 — around 23 20-step tasks. Below that volume, pay-as-you-go wins because a quiet month costs nothing. Above it, the subscription is worthwhile — until you hit the plan limits and have to wait for the window to reset.

Paying from Brazil — Pix, no international card

This is usually the real obstacle, not the price. A subscription purchased on the website accepts only credit or debit cards (Anthropic help center, checked on October 1, 2026); in the mobile app, billing is handled by the App Store or Google Play. The Kunavo API-key route does not depend on a card: checkout is processed by Stripe, and Pix appears as a payment method for customers paying from Brazil, with the amount shown in reais and immediate settlement. International cards (Visa, Mastercard, American Express), Apple Pay, and Google Pay remain available for those who prefer them.

Per-token rates are set in USD: with Pix, Stripe shows the conversion to reais at the time of payment; with a card, conversion follows the issuer’s exchange rate (check the IOF and international transaction fee in your banking app) — step-by-step instructions for paying with Pix are in the guide to paying for the API with Pix. The wallet is prepaid — a $10 minimum top-up, no subscription, no automatic renewal, and a balance that never expires. Based on the calculation above, $10 via Pix covers about 12 20-step agentic tasks. The full Brazil context (São Paulo latency, LGPD, invoices) is in the Claude API in Brazil guide.

How to point Claude Code to Kunavo

Claude Code reads the endpoint and credentials from environment variables—six lines, with no proxy or plugin:

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 are not optional: Claude Code's default model and the opus and sonnet aliases follow Anthropic's newest models. 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 subagents with model: sonnet return 404. The opus alias is fixed to Claude Opus 5.5 (claude-opus-5-5, $2.80 / $14.00 per 1M tokens on Kunavo), which requires Claude Code 2.1.280 or later—on an earlier version, run claude update. While ANTHROPIC_AUTH_TOKEN is set, Claude Code uses the key instead of the saved login: plan limits no longer apply, and usage is charged to the key balance. Remove the variable to return to the subscription; you can switch between the two routes by shell. Create the key in the dashboard; the complete walkthrough is in the Claude API key guide.

Calculating the cost of a real session

Run /usage inside Claude Code to see the session’s four numbers (new input, output, cache read, and cache write), then plug them into this function:

# Quanto custou uma sessão do Claude Code, a partir dos tokens dela.
# Rode /usage dentro do Claude Code para pegar os quatro números reais.
RATES = {  # USD por 1M de tokens na Kunavo: (input, output, leitura de cache / input)
    "claude-haiku-4-5":  (0.70, 3.50, 0.1),
    "claude-sonnet-4-6": (2.10, 10.50, 0.1),
    "claude-sonnet-5":   (1.40, 7.00, 0.1),
    "claude-opus-5":     (3.50, 17.50, 0.1),
    "claude-opus-5-5":   (2.80, 14.00, 0.05),
}

def custo_sessao(model, inp, out, cache_read=0, cache_write=0):
    i, o, r = RATES[model]
    return (
        inp / 1e6 * i                   # input novo
        + out / 1e6 * o                 # output, incluindo tokens de raciocínio
        + cache_read / 1e6 * i * r      # leitura de cache: fração r da tarifa de input
        + cache_write / 1e6 * i * 1.25  # escrita de cache de 5 min: 125%
    )

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

Notice the weight of cache reads: in a long session they dominate the token count, and they are precisely what costs only 10% of the input rate on Sonnet 5 (5% on Opus 5.5). That is why prompt caching, rather than model choice, is the single biggest cost lever in Claude Code.

How to lower your bill

1. Let caching do its work.Cached input costs 10% of the rate on Sonnet 5 (5% on Opus 5.5), and Claude Code rereads the entire conversation at every step (how it works).
2. Clear the context between unrelated tasks. An endless session resends the entire history at every step; shorter context costs less.
3. Choose a model to fit the task. Leaving the largest model as the default is the most common and costly mistake.
4. Route internal calls to Haiku with ANTHROPIC_DEFAULT_HAIKU_MODEL — they’re frequent and simple.

Haven’t installed it yet? The walkthrough is in how to install Claude Code. If you’re comparing it with OpenAI’s agent, the same calculation for Codex is in how much Codex costs.

Frequently asked questions