Claude Pro usage limits aren’t a fixed monthly message count; they use rolling windows: a shorter session window (the “5-hour limit” people talk about) plus a longer weekly cap, with chat and Claude Code sharing the same window. How much you can do in a window depends on the model and conversation length—Anthropic doesn’t publish a fixed number and adjusts the limits. When you use up a window, the plan stops serving you until it resets; paying more within the same plan won’t buy more capacity. You then have only three options: wait, upgrade, or continue with a per-token API key. The same 20-step task costs about $0.868 with Claude Sonnet 5.
Most articles about Claude limits stop at explaining “how they work.” This page goes on to calculate how much each of the three options costs after you hit the limit, and how to keep your subscription while using a key only during blocked periods.
This page deliberately doesn’t list allowance figures. Anthropic adjusts them, and figures copied by third-party sites quietly go out of date. Check the Anthropic Help Center for current figures. Plan prices are based on claude.com/pricing as of September 2026; Kunavo rates were checked on October 5, 2026 and are reflected in this site’s model catalog.
How Claude limits work
| Subscription (Pro / Max) | API key | |
|---|---|---|
| Usage metering | Rolling windows: session window plus weekly cap | No window; deducts from prepaid balance |
| When you hit the limit | Stops serving you until the window resets | Keeps serving you until the balance runs out |
| Months with no activity | Monthly fee continues | $0 |
| A heavy-use week | Capped — you can’t buy more capacity within the same plan | Pay for what you use |
| Budget pattern | Fixed and predictable | Variable, with a limit you set |
Two features determine the practical impact. First, chat and Claude Code share the same window: a round of discussion in chat in the afternoon leaves less capacity for coding that evening. Second, the limit doesn’t hit evenly; it tends to hit during the hours when you get the most done. The day you’re capped is often your most valuable day. That cost isn’t a price, so it won’t appear in any pricing comparison table.
Three options after you hit the limit
| Option | Cost | Best for |
|---|---|---|
| Wait for the window to reset | $0, but work stops | There’s no deadline today; you can wait a few hours |
| Upgrade to a higher-tier plan | Monthly price difference (Pro → Max is $80 as of September 2026) | You hit the limit several times a week and consistently need more capacity |
| Continue with an API key | Per token: a 20-step task on Sonnet 5 costs about $0.868 | You hit the limit only a few times a month and need to deliver today |
The third option doesn’t require canceling your subscription. Set two environment variables to switch over; we’ll provide the setup below. First, let’s break down its cost.
What the same work costs per token
The unit used across all Claude Code pricing pages on this site is the agentic step: one back-and-forth with the model within a task, with about 25,000 input and 1,200 output tokens. A task is 20 steps. Rates are loaded live from the model catalog:
| Model | Anthropic official (input / output) | Kunavo (input / output) | Price difference | One step | One 20-step task |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | $1.00 / $5.00 | $0.70 / $3.50 | Save approximately 30% | $0.022 | $0.434 |
| Claude Sonnet 5 | $2.00 / $10.00 | $1.40 / $7.00 | Save approximately 30% | $0.043 | $0.868 |
| Claude Opus 5 | $5.00 / $25.00 | $3.50 / $17.50 | Save approximately 30% | $0.109 | $2.17 |
For small background tasks, use Haiku 4.5, which costs just $0.022 per step.
The table above doesn’t include caching. Cache reads are charged at 10% of the input rate, and writes at 1.25 times the input rate. Claude Code resends the entire conversation at every step, so the input side costs much less when the cache is hit. See Claude prompt caching (in English) for details.
Upgrade or top up with a key? One calculation
First, look at the whole plan. As of September 2026, Pro costs $20 per month, equivalent to about 23 20-step tasks’ worth of Sonnet 5 API usage; Max starts at $100 per month, equivalent to about 115 tasks. If you use it every day, do more than this number of tasks a month, and stay within the limit most of the time, a subscription costs less than per-token billing—for heavy use, the subscription is the cheaper option. The API key is useful for the hours when you’re blocked.
Now consider when you’re blocked. Suppose upgrading means you no longer hit the limit. The monthly price difference between Pro and Max is $80. Divide that by the cost of topping up with a key each time you’re blocked to find how many times a month you’d need to hit the limit for upgrading to cost less:
| Still need to finish while blocked | Cost of one top-up with a Sonnet 5 key | How many times you need to hit the limit in a month for upgrading to cost less |
|---|---|---|
| 1 20-step tasks | $0.868 | About 92 times |
| 3 20-step tasks | $2.60 | About 31 times |
| 10 20-step tasks | $8.68 | About 9 times |
This table only compares costs and doesn’t account for time spent waiting. The conclusion matches intuition, but now it has numbers: if you hit the limit several times a week and need to top up a lot each time, upgrading is worth it; if you hit it once or twice a month, topping up with a key is cheapest. Full Claude plan and API pricing is in Claude pricing; the Claude Code break-even calculation is in Claude Code pricing.
Keep your subscription and use a key only when you’re blocked
A subscription and an API key are separate credentials, so for most heavy users the answer isn’t “switch” but “keep both”: use the subscription for chat and regular sessions, and the key for the hours when the plan stops serving you.
# 保留訂閱，只在被擋下的時段改用金鑰。
# 這兩個變數存在時，Claude Code 用金鑰，而不是你的登入身分。
export ANTHROPIC_BASE_URL=https://api.kunavo.com # origin，不要加 /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
# 窗口重置後拿掉變數，就回到訂閱：同一個安裝、同一個專案。
# unset ANTHROPIC_BASE_URL ANTHROPIC_AUTH_TOKEN ANTHROPIC_MODEL ANTHROPIC_DEFAULT_OPUS_MODEL ANTHROPIC_DEFAULT_SONNET_MODEL ANTHROPIC_DEFAULT_HAIKU_MODELWhen the variable is present, Claude Code bills by API key and the subscription is completely unaffected; remove the variable to return to the subscription. The same installation and project remain in place, with nothing to migrate or restore. ANTHROPIC_BASE_URL is the origin; do not add /v1. The model lines cannot be omitted: Claude Code's default model and the opus alias both point to the latest Opus, so if Kunavo has not listed it yet, the first request will receive a 404. Here, pin the opus alias to Opus 5.5 (claude-opus-5-5); Claude Code v2.1.280 or later is required, and older versions should first run claude update. The sonnet alias defaults to Sonnet 5.5, which Kunavo does not provide, so without the ANTHROPIC_DEFAULT_SONNET_MODEL line, /model sonnet and opusplan during execution outside plan mode, as well as subagents assigned model: sonnet, will receive a 404. If what is blocked is chat rather than Claude Code, the same key can also connect to a chat client that supports an OpenAI-compatible or Anthropic-compatible API, with the same per-token billing. See the Claude Code integration documentation and Claude Code API key setup for configuration details (English).
Before giving a key to an agent
The key removes the plan’s cap—that’s why you’re using it—but know what replaces that cap. Kunavo has two layers. Prepaid balance is the outer limit: usage is deducted from the amount you’ve added, rather than charged to the card on file, so an agent can spend at most what’s in the account. Within that, you can set a monthly spending limit for each key. It’s checked before every request, returns 402 when reached, and resets at the start of each month. You can also set an IP allowlist. Give each unattended agent its own key; keys can be revoked individually, limiting any issue to that one agent. Details are in the billing documentation.
To be candid
Claude requests through the gateway use shared capacity: there’s no dedicated quota or contractually guaranteed SLA. For a scale that needs guaranteed capacity or an SLA, contract directly with Anthropic. And as the calculations above show, a subscription is already cheaper for people who use it daily and stay within the limit most of the time—use a key to fill the gaps, not as a replacement.
For payments from Taiwan: Kunavo uses Stripe prepaid billing and supports cards including JCB (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, and Google Pay, with a minimum top-up of $10 and no balance expiration; Taiwan has no local payment channel, and JKoPay and LINE Pay are not on the available list. The English original of this series is Claude Pro and Max limits.
Frequently asked questions
What are Claude Pro usage limits?
Claude paid plans use rolling windows rather than a fixed monthly allowance: a shorter session window plus a longer weekly cap, shared by chat and Claude Code. How much you can do in a window depends on the model and conversation length. Anthropic doesn’t publish a fixed message count and adjusts limits, so its own Help Center is the reliable source. The pattern remains the same: when you use up a window, the plan stops serving you until it resets, and paying more within the same plan won’t change that.
What is Claude’s 5-hour limit?
It’s the rolling session window on paid plans: usage is measured over a period of several hours, not by calendar day. A short, intensive session can use up a whole window, while a light day might not reach the limit at all. In practice, the cap tends to hit right when you’re most productive—which makes the choice between a subscription and an API key a question of your usage pattern, not your total usage.
What happens when you hit Claude’s usage limit?
Your session stops until the window resets; there’s no way to pay to get past the limit within the same plan. There are three real options: wait, upgrade to a higher tier, or switch to an API key in the same client so work can continue with per-token billing. Claude Code supports the third option natively: it reads ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN. Set them to use a key instead of your plan; remove them to switch back to your plan.
How much does it cost to switch to the Claude API when I run out of credits?
Per agentic step (25,000 input / 1,200 output tokens), Claude Sonnet 5 on Kunavo costs about $0.043, Claude Opus 5 about $0.109, and Claude Haiku 4.5 about $0.022. A 20-step task on Sonnet 5 costs about $0.868. These figures don’t include caching; cache reads are charged at 10% of the input rate, so actual sessions cost less when the cache is hit.
Is upgrading to Claude Max worth it?
It depends on how often you hit the limit, not your total usage. As of September 2026, Pro costs $20 per month and Max starts at $100 per month (see claude.com/pricing), a difference of $80. If you need to finish 3 tasks of 20 steps each when you’re blocked, topping up with a Sonnet 5 key costs about $2.60. You’d need to hit the limit about 31 times or more per month for upgrading to cost less. If you hit it several times a week and your work has to stop, upgrading is worth it; if you hit it once or twice a month, topping up with a key is cheapest.
Can I use a Claude subscription and an API key at the same time?
Yes, and this is the setup most heavy users end up with. They’re separate credentials, not mutually exclusive: the subscription covers chat and regular sessions, while the API key covers the hours when the plan stops serving you. In Claude Code, setting the two environment variables pauses subscription use only in that shell. Your plan, installation, and project remain unchanged; remove the variables to switch back immediately. The two accounts are billed separately, with no offsets between them.
Do API keys also have usage limits?
There’s no rolling window that stops you mid-session like a plan’s limit, but you can set your own limits. Your prepaid balance is the outer cap because Kunavo deducts from your prepaid balance instead of charging the card on file. Within that, you can set a monthly spending limit for each key. It’s checked before every request; requests return 402 once the limit is reached, and the limit resets at the start of each month. Unlike a plan’s allowance, you choose the amount and can raise it immediately in the dashboard when needed.