Rate limits exist for two reasons
Every API has rate limits. For Claude, they serve two purposes: protecting Anthropic's infrastructure from abuse, and ensuring fair access across all customers. If you're hitting 429s, understanding the model helps you work around them.
Anthropic's native rate limits
Anthropic enforces limits at several levels:
- Requests per minute (RPM) — how many API calls you can make per 60 seconds
- Tokens per minute (TPM) — how many tokens (input + output) per 60 seconds
- Tokens per day (TPD) — a daily ceiling on total consumption
The exact numbers depend on your tier. Free tier users have tight limits. Tier 4 and above (which requires significant spend) get much higher ceilings.
The core problem: these reset on rolling-minute and calendar-day boundaries. If you're running Claude Code intensively, you'll hit the day limit and be locked out until midnight UTC.
How rolling windows are different
Antex uses a different model: two rolling windows (5 hours and 7 days) that reset continuously, not at midnight.
A rolling 5h window means: your usage budget covers any 5-hour period. If you used 50% of your budget 3 hours ago, 3 hours from now that usage falls off the window and you get it back. There's no cliff at midnight.
This matters practically: you can run a heavy Claude Code session in the morning, take a break, and come back in the afternoon with recovered capacity — without waiting until the next day.
What a 429 response looks like
A standard Anthropic 429:
{
"type": "error",
"error": {
"type": "rate_limit_error",
"message": "Rate limit exceeded..."
}
}On Antex, 429s tell you which window is exhausted and when it clears:
{
"type": "error",
"error": {
"type": "rate_limit_error",
"message": "5h window exhausted. Resets at 14:32 UTC."
}
}Claude Code displays this as a rate limit error with the reset time.
How to avoid hitting limits
1. Check your dashboard before heavy sessions — the percentage bars show remaining capacity in both windows.
2. Use Sonnet, not Opus, for most tasks — Opus uses significantly more capacity per request.
3. Avoid 1M context requests for small tasks — large context = large capacity hit.
4. Spread work across the day — rolling windows mean you recover as you go.
Concurrency limits
Each plan also has a concurrent request limit — how many requests can be in-flight simultaneously. Claude Code normally uses 1 at a time, but if you're running multiple sessions (e.g. multiple terminal windows), each counts. The Pro plan allows 5 concurrent; Ultra allows 20.