Rate limits & concurrency

The platform enforces multiple limits for fairness and stability. Hitting one returns 429 rate_limit_exceeded.

Limit types

Hitting a limit

• Returns 429 rate_limit_exceeded; the retry-after header suggests how long to wait.
• Not billed: rate-limited requests incur no charge.

Pre-authorization

To prevent overspending, the platform conservatively reserves an amount before each request, settles against actual usage afterward, and releases the difference. Reservations are not final balance — only actual usage counts.

Recommendations

• Implement exponential backoff and honor retry-after.
• Contact support to raise limits for high-concurrency workloads.