Rate Limits

The two voice endpoints have their own fixed, per-API-key rate limits — independent of the general Public API rate limits.

Limits

Limit	Value
`GET /voice-agents`	100 requests/minute
`POST /voice-agents/:agentId/sessions`	10 requests/minute
Concurrent sessions per key	5
Maximum session duration	3600 seconds (1 hour)

A session still running at the 3600-second mark is closed automatically; session_state is set to completed with a timeout reason.

Every response from either endpoint carries these headers:

Header	Description	Example
`X-RateLimit-Limit`	The endpoint's limit, requests/minute	`100`
`X-RateLimit-Remaining`	Requests left in the current 60-second window	`37`
`X-RateLimit-Reset`	Unix timestamp (seconds) when the current window ends	`1732550460`
`X-RateLimit-Retry-After`	Seconds until the window resets — only present on `429` responses	`23`

A 429 also carries a plain Retry-After header with the same value as X-RateLimit-Retry-After.

Cache the discovery response. Your list of voice-enabled agents rarely changes — call GET /voice-agents once and cache the result instead of calling it before every session.
Use exponential backoff on 429. Read retry_after from the response body (or Retry-After header) and back off at least that long; add jitter if you're retrying multiple sessions at once.
Monitor X-RateLimit-Remaining. Slow down proactively when it gets low instead of waiting for a 429.
Reuse sessions where possible. Each session start counts against both the 10 req/min limit and the 5-concurrent-session cap — end sessions promptly (see Session Management) so the slot is free for the next call.