The limits are not the interesting part. What you do at the limit is. Every guide to Claude's plans explains that usage is metered in rolling windows and that the window resets. Almost none of them answers the question you actually have when the window closes at four in the afternoon: is it cheaper to wait, to upgrade, or to keep working on a per-token key?
That is one division, and this page does it. It deliberately does not reprint Anthropic's plan prices or quota numbers — those move, and a page quoting a stale quota is worse than a page quoting none. Anthropic's own help centre is the source for what the limits currently are; what follows is what they cost you.
How the limits behave
| Subscription | Per-token key | |
|---|---|---|
| Metering | Rolling windows — a session window and a longer weekly cap | None. You spend down a balance |
| At the limit | Stops serving until the window resets | Keeps serving until the balance runs out |
| A quiet month | Charged in full | Charged nothing |
| A very heavy week | Capped — you cannot buy through it in-tier | Bills what it bills |
| Budget shape | Fixed and predictable | Variable, capped only by what you set |
The row that decides most cases is the fourth. A plan caps your worst day, and your worst day is usually your most valuable one. That is a real cost, and it never shows up in a price comparison because it is not a price.
The break-even
# What a plan has to save you to pay for itself.
#
# plan_price_per_month / cost_per_task = tasks/month to break even
#
# A 20-step task on Claude Sonnet 4.6 at Kunavo rates costs $0.744:
#
# $20/month -> 27 tasks
# $100/month -> 134 tasks
# $200/month -> 269 tasks
#
# Divide by ~22 working days for a per-day figure. If you are not
# clearing that most days, the plan is buying you headroom you
# are not using — and the headroom is exactly what the limits cap.A step here is 25,000 input and 1,200 output tokens — one agentic round trip in a coding client, sized the same way the Claude Code page and the Codex page size it, so all three are comparable. Rates are read from the catalog at render time:
| Model | Kunavo, per 1M in / out | Anthropic list | Per 20-step task |
|---|---|---|---|
| Claude Haiku 4.5 | $0.40 / $2.00 | $1.00 / $5.00 | $0.248 |
| Claude Sonnet 4.6 | $1.20 / $6.00 | $3.00 / $15.00 | $0.744 |
| Claude Opus 5 | $2.00 / $10.00 | $5.00 / $25.00 | $1.24 |
Keeping the plan and adding a key
These are separate credentials, so the honest answer for most heavy users is not “switch” — it is both. Keep the plan for the chat app and ordinary sessions; put a key behind the client for the hours the plan has stopped serving.
# Keep the plan. Add a key for the hours the plan has cut you off.
#
# Claude Code reads these two variables and stops using the subscription
# while they are set — so this is a switch, not a migration.
export ANTHROPIC_BASE_URL=https://api.kunavo.com
export ANTHROPIC_AUTH_TOKEN=sk-kn-...
export ANTHROPIC_MODEL=claude-sonnet-4-6
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5
# Unset them and you are back on the plan, same install, same session.While those variables are set, Claude Code bills the key and the plan is untouched. Unset them and you are back, same install, same project, same session history. There is no migration and nothing to undo — which is what makes the break-even above a live decision rather than a one-way door. Setup detail is on getting an API key for Claude Code, and what is free and what is not covers the ground before either.
Before you hand a key to an agent
A per-token key removes the plan's ceiling, which is the point of it — but it is worth knowing what replaces that ceiling. On Kunavo two things do. The prepaid balance is the outer bound: billing draws down an amount you topped up rather than charging a card, so the most an agent can spend without you is the balance sitting there. Inside that, a key can carry its own monthly spend limit, checked before each request and reset at the start of the calendar month — which is the one to set deliberately if the agent runs unattended or with permission prompts turned off, and keep a separate key per agent — a key can be revoked on its own, which makes the blast radius one agent rather than the account.
FAQ
What are the Claude Pro usage limits?
Claude's paid plans meter usage in rolling windows rather than a fixed monthly credit — a short session window plus a longer weekly cap — and the amount of work that fits inside a window depends on the model and on how long your conversations are, not on a published request count. Because Anthropic adjusts these numbers, the reliable source is Anthropic's own help centre rather than any third-party page, this one included. What is stable is the shape: when a window is exhausted the plan stops serving until it resets, and no amount of paying more inside the same tier changes that.
What happens when you hit the Claude usage limit?
The session stops until the window resets — you cannot buy your way through it inside the plan. The three real options are to wait, to move to a higher tier, or to put an API key in front of the same client so work continues billed per token. Claude Code supports the third natively: it reads ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, and while those are set it uses the key instead of the plan, so it is a switch you can flip in one shell and flip back.
Is the Claude Max plan worth it?
Divide the plan price by what one task costs per token and you get the number of tasks the plan must save you. A twenty-step coding task on Claude Sonnet 4.6 is about $0.744 at Kunavo rates, so a $100 plan breaks even near 134 tasks a month and a $200 plan near 269 — roughly 12 substantial tasks every working day for the larger tier. If your usage clears that most days, the flat fee wins and the predictability is a bonus. If it is bursty, per-token wins outright, because the weeks you do not code cost nothing.
Can I use Claude Pro and an API key at the same time?
Yes, and this is the setup most heavy users land on. They are separate credentials, not an either/or on your account: the plan serves the chat app and everyday sessions, and an API key serves the client during the hours the plan has cut you off. In Claude Code, setting the two environment variables parks the subscription for that shell only — nothing about the plan, the install or the session history changes, and unsetting them puts you straight back.
What is the Claude 5-hour limit?
It is the rolling session window on Anthropic's paid plans: usage is metered over a few hours rather than per calendar day, so heavy work in a short burst can exhaust a window while a light day never touches it. The practical consequence is that the limit binds precisely when you are most productive, which is why the plan-versus-key question is really a question about the shape of your usage rather than its total.
Which Claude model should I route to when I switch to a key?
Match the model to the step. Claude Sonnet 4.6 at $1.20 / $6.00 per 1M tokens is the working default; Claude Opus 5 at $2.00 / $10.00 is worth it for architecture-level changes and wasteful for routine edits; Claude Haiku 4.5 at $0.40 / $2.00 should carry the background sub-tasks a coding client fires on its own. Setting ANTHROPIC_DEFAULT_HAIKU_MODEL is one line for a permanent reduction, because those background calls are frequent and invisible.
Do API keys have usage limits too?
Not in the plan sense — there is no rolling window that stops you mid-session — but you can impose your own. The prepaid balance is the outer ceiling, because Kunavo bills down a balance rather than charging a card on file. Inside that, each key can carry a monthly spend limit you set yourself, checked before a request runs and reset at the start of each calendar month. The difference from a plan's quota is who chooses the number and how quickly it changes: you set it, and you can raise it in the dashboard the moment it gets in your way.