LiveKit Inference STT returns MaxConcurrentGatewaySTT with 0 usage despite 5-connection quota

[Antony_Esquivel]

I’m encountering what appears to be an inconsistency between my LiveKit Cloud project’s displayed STT quota and the quota enforced by the LiveKit Inference gateway.

My Quota and Limits dashboard shows:

• Model: Deepgram Nova-3 (Monolingual)
• Type: STT
• Limit: 5 concurrent connections
• Peak usage (past 7 days): 1/5

However, attempts to establish a LiveKit Inference STT connection are currently rejected with HTTP 429.

I captured the JSON error returned by the STT WebSocket upgrade:

type: inference_quota_exceeded
error: STT connection limit exceeded

category: MaxConcurrentGatewaySTT
remaining_limit: 0
current_usage: 0
status: QuotaStatusExceeded

The important inconsistency is that the gateway reports current_usage: 0 and remaining_limit: 0, while the project dashboard shows an STT concurrency limit of 5.

The integration was previously working. My September billing/usage page currently shows 216 minutes of Deepgram Nova-3 STT usage, so the model has successfully been used by this project before.

I have verified the client/agent audio path independently: the microphone track reaches the LiveKit agent, RTP packets are received, decoded PCM audio reaches the agent’s STT pipeline, and the failure occurs specifically when the LiveKit Inference STT WebSocket connection is established.

Could someone from LiveKit please check whether the effective MaxConcurrentGatewaySTT quota for this project is incorrectly set to zero or whether another account/project-level restriction is causing the gateway to return this quota error?

I can provide the affected project ID and gateway session ID privately if needed.

Thank you.

[abidullahcs.uk (Abid)]

@Antony_Esquivel Worth adding the credit math to your post, because it rules out the answer you’ll otherwise get first.

The Build plan includes $2.50/month of inference credits, so “you ran out of credits” is the obvious reply. Your own September numbers say otherwise:

Item	Usage	Rate	Cost
Nova-3 STT	216 min	$0.288/hr	$1.04
Inworld TTS	15,812 chars	$25/M chars	$0.40
Gemma 4 31B - Input	348,719 tokens	$0.40/M tokens	$0.14
Gemma 4 31B - Output	8,281 tokens	$1.20/M tokens	$0.01
		Total	~$1.58

Roughly 63% used, so there’s headroom.

The second point is worth stating plainly: current_usage: 0 with remaining_limit: 0 means the effective limit is zero. If you were genuinely saturating 5 connections it would read current_usage: 5. So the gateway is applying a limit of 0 where your dashboard shows 5.

Both point the same way, and neither is something you can fix from your side. Adding the math should stop it being closed as expected quota behaviour.