Every call to our deployed agent connected and then went silent. Agent logs show the Inference gateway rejecting every TTS and LLM request:

Telephony
1

Written for: LiveKit support engineers, to paste into a ticket.

Subject: Inference gateway returned 429 quota_exceeded (max_allowed 0) while credit balances were “temporarily unavailable”; agent went silent on live calls

Hi LiveKit team,

Project: sol-reception (p_37abpdaoh6k), Build plan, us-east. Agent CA_BBFZsTiUbuiw.

On Oct 24, 2026 around 08:30 UTC, every call to our deployed agent connected and then went silent. Agent logs show the Inference gateway rejecting every TTS and LLM request:

failed to synthesize speech: Invalid response status (429 Too Many Requests)
LLM token credit quota exceeded, category: MaxGatewayCredits,
remaining_limit: 0, current_usage: 0, max_allowed: 0,
status: QuotaStatusExceeded, free_tier: true
project_id: p_37abpdaoh6k
example session_id: 86c28aea-2e9b-42f0-922d-42f17527524e
example room: sol-reception-call-_+1619XXXXXXX_zG78kcMFRs4o

Models affected: cartesia/sonic-3 (TTS) and google/gemma-4-31b-it (LLM).

At the same time the project Billing page showed “Credit balances are temporarily unavailable.” October usage is well inside the Build plan’s included inference credits (about 940K Gemma tokens and 44K Sonic characters, all billed at $0), and the Usage and limits page shows no rate limit hit at the time of the failures. The agent had been running normally through the same gateway from Oct 19 to 23.

It looks like the gateway’s credit check failed closed while the balance service was down, reporting a zero allowance for a project that had credits. Three questions:

  1. Can you confirm that is what happened, and whether it is resolved?
  2. Is failing closed on a balance-service outage the intended behavior? For a phone agent it means callers hear silence. A fail-open with a short grace window, or at least a distinct error code so clients can fall back, would help a lot.
  3. Is there a status signal or webhook we can watch for this condition?

We have since moved TTS and LLM to direct provider calls, so we are not blocked, but we would like to understand the failure before relying on the gateway again.

Thanks,
Lainie
LNL Group / Sol Reception

2

Hi, apologies for the confusing messages, as my colleague has said in other threads, we are working on improving this.

In your case if I look at your usage page at https://cloud.livekit.io/projects/p_/settings/billing you have used the equivalent of about $2.21 of Cartesia TTS and $0.24 of Gemma, which puts you around the $2.50 Inference limit on build plans. Note that the limit is cumulative across all Inference providers.