Paid Claude plans measure usage in rolling windows—a short session window and a longer weekly limit—and the session stops when you hit the limit. Nearly every page on this topic explains that and stops there. This page goes further: what the work costs per token, where the break-even point lies compared with a subscription, and how to keep your subscription while adding a key only for the hours when you’re blocked.
This page deliberately omits specific allowance figures. Anthropic adjusts them, and third-party sites that repeat them can quietly become inaccurate. Anthropic’s Help Center is the only reliable source for current figures. The structure—and the arithmetic behind it—remains stable. Rates verified on October 3, 2026.
How the limits work
| Subscription (Pro / Max / Team) | API key | |
|---|---|---|
| Billing | Flat monthly fee | Per token, no monthly fee |
| Limits | Rolling windows—a session window and a weekly limit | No window; prepaid balance and a key limit you set |
| At the limit | Session stops until reset | Continues until the balance runs out |
| Quiet month | Full monthly price | $0 |
Two properties make the practical difference. First, Claude chat and Claude Code share the same window—an afternoon of discussion in chat reduces what the coding session can use that evening. Second, the limit doesn’t apply evenly; it kicks in during the very hours when you’re getting the most done.
The break-even point—one division
The break-even point depends on the number of steps, not the number of workdays. An agentic step is one exchange with the model within a task, roughly 25,000 input tokens and 1,200 output tokens—the same unit used in Claude pricing.
| Model | Per step | 20 steps (one task) | 100 steps |
|---|---|---|---|
| Claude Haiku 4.5 | $0,0217 | $0,43 | $2,17 |
| Claude Sonnet 5 | $0,0434 | $0,87 | $4,34 |
| Claude Opus 5 | $0,11 | $2,17 | $10,85 |
This gives us:
- $20 (Pro) ≈ 460 steps per month ≈ about 23 tasks with 20 steps each
- $100 (Max, entry tier) ≈ about 2.300 steps
If your usage falls below that or is irregular, pay-as-you-go costs less—a quiet month then costs nothing. If you regularly exceed it, the flat fee is better until you hit the usage limits. And right in that gap—regular enough to need a subscription, but locked out several times a month—is a third option.
Keep the subscription and add a key
You don’t have to choose. Claude Code uses an API key while the environment variables are set, then falls back to your subscription once they’re removed.
# Das Abo behalten und für die gesperrten Stunden einen Key dazulegen.
# Solange diese Variablen gesetzt sind, nutzt Claude Code den Key statt des Logins.
export ANTHROPIC_BASE_URL=https://api.kunavo.com
export ANTHROPIC_AUTH_TOKEN=sk-kn-...
# Modelle festlegen: Standardmodell und die Aliasse opus und sonnet folgen bei
# Claude Code den neuesten Anthropic-Modellen. Der Alias sonnet fragt Sonnet 5.5
# an, das Kunavo nicht anbietet – ohne Festlegung enden /model sonnet, die
# Ausführungsphase von opusplan und Sonnet-Subagenten mit 404.
# Opus 5.5 braucht Claude Code ab v2.1.280 (ältere Versionen: claude update).
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
# Variablen entfernen -> zurück aufs Abo.This means you pay for the subscription during the hours when it works and for tokens only during the hours when it locks you out. The two forms of billing are separate and don’t offset each other: a subscription doesn’t reduce the API bill, and a key doesn’t increase the subscription allowance.
The same models cost 30% less than Anthropic’s list prices on Kunavo . See Claude pricing for the full table and the quickstart guide for setup instructions.
Before you give an agent a key
An API key has no rolling window—that’s the point, and also the risk. An agent stuck in a loop won’t be stopped at 4 p.m.; it will keep running. Two limits help prevent that:
- Prepaid balance is the outer ceiling. Kunavo deducts from your balance instead of charging a saved card, so it can’t charge you for more than you’ve already deposited.
- A spending limit for each key, which you set yourself: it’s checked before every request and resets at the start of the month. Unlike a plan allowance, you choose the amount and can raise it immediately.
For an unattended agent, this is the real answer to the limit question: instead of a window that blocks you, you set your own cap. Details about keys and billing are in the billing documentation.
For an overview of Claude’s total cost—subscription, API, and Claude Code side by side—see Claude pricing. Find per-model rates in the model overview.
Frequently asked questions
What are the Claude Pro limits?
Paid Claude plans measure usage in rolling windows instead of a fixed monthly allowance: a short session window plus a longer weekly limit. How much work fits into a window depends on the model and the length of your conversations, not on a published number of requests. Anthropic adjusts these figures, so its own Help Center—not a third-party site, including this one—is the reliable source. The structure is stable: once a window is exhausted, the plan won’t provide more usage until it resets.
What is Claude’s 5-hour limit?
The rolling session window for paid plans: usage is measured over a period of several hours, not by calendar day. Intensive work in a short time can exhaust the window, while a quiet day may never reach the limit. In practice, this means the limit can kick in precisely when you’re most productive—which is why the choice between a subscription and an API key is really about the shape of your usage, not its total amount.
Is Claude Max worth it?
It depends on how often you hit the limit, not how much you use. If you hit it several times a week and your work comes to a stop, the extra $80 over Pro pays for itself through the waiting time you save. If you hit it once a month, it’s cheaper to cover those hours with an API key: an agentic step (25,000 input tokens and 1,200 output tokens) costs about $0,0434 with Claude Sonnet 5 on Kunavo, so one blocked afternoon costs cents instead of a monthly plan upgrade.
What does a Claude subscription cost compared with an API key?
The comparison is based on the number of steps, not workdays. One agentic step costs around $0,0434 on Kunavo with Claude Sonnet 5. A $20 Pro subscription therefore corresponds to about 460 steps per month, or roughly 23 tasks of 20 steps; $100 (Max, entry level) corresponds to about 2.300 steps. Below that level, pay-as-you-go is cheaper—a quiet month costs nothing. If you regularly exceed it, the flat-rate plan wins until the usage limits take effect.
Can I keep my subscription and still use an API key?
Yes, and that’s usually the right answer. Claude Code uses an API key while the relevant environment variables are set, then falls back to your subscription once they’re removed. You pay for the subscription during the hours when it works and for tokens only during the hours when it locks you out. The two forms of billing are separate and don’t offset each other.
Does an API key also have a usage limit?
Not in the sense of a plan limit—there’s no rolling window that stops a session partway through. You can, however, set a limit yourself. Prepaid credit is the outer boundary because Kunavo deducts from your balance instead of charging a saved card. Each key can also have its own monthly spending limit, which is checked before every request and resets at the start of the month. Unlike a plan allowance, you choose the amount and can raise it immediately in the dashboard if it gets in the way.
What happens when I hit the Claude limit?
The session stops until the window resets, and paying more won’t change that within the same tier—paying more means subscribing to a higher tier. You have three options: wait, switch to the next tier, or set an API key for those hours and then switch back to the subscription. The third option is the only one that doesn’t require a monthly commitment.