Records
Usage
What every model used on this deployment, and what it cost, by who bills it. How each figure is worked out is under How to read this, at the end.
By provider
Google · Gemini API
$2.3322 sessions · free tier — the cost is the list price, not charged
- Gemini 3.8 Live$2.17
- Gemini 3.8 Live Extended Thinking$0.161
On the free tier Google may use what callers say, and people may read it: set up billing in AI Studio before a real guest calls. Its limits per model are in AI Studio, under Rate Limit.
Qwen · Model Studio
$0.00Nothing used in this period.
Free quota, counted from what this deployment recorded — a lower bound. Model Studio's console has the real balance.
Estimated cost per day
Bali days · point at or tab to a day for what it went onShow the numbers
| Day | Estimated cost | Spent on |
|---|---|---|
| 2026-10-12 | $0.00 | — |
| 2026-10-13 | $1.57 | Brain $1.57 |
| 2026-10-14 | $0.101 | Brain $0.101 |
| 2026-10-15 | $0.00 | — |
| 2026-10-16 | $0.00 | — |
| 2026-10-17 | $0.00 | — |
| 2026-10-18 | $0.00 | — |
| 2026-10-19 | $0.184 | Brain $0.184 |
| 2026-10-20 | $0.00 | — |
| 2026-10-21 | $0.00 | — |
| 2026-10-22 | $0.00 | — |
| 2026-10-23 | $0.00 | — |
| 2026-10-24 | $0.00 | — |
| 2026-10-25 | $0.00 | — |
| 2026-10-26 | $0.479 | Brain $0.479 |
| 2026-10-27 | $0.00 | — |
| 2026-10-28 | $0.00 | — |
| 2026-10-29 | $0.00 | — |
| 2026-10-30 | $0.00 | — |
| 2026-10-31 | $0.00 | — |
| 2026-11-01 | $0.00 | — |
| 2026-11-02 | $0.00 | — |
| 2026-11-03 | $0.00 | — |
| 2026-11-04 | $0.00 | — |
| 2026-11-05 | $0.00 | — |
| 2026-11-06 | $0.00 | — |
| 2026-11-07 | $0.00 | — |
| 2026-11-08 | $0.00 | — |
| 2026-11-09 | $0.00 | — |
| 2026-11-10 | $0.00 | — |
By model
| Model | Part | Sessions | Usage | List price | Est. cost | Share | Latency |
|---|---|---|---|---|---|---|---|
| Gemini 3.8 Live | Brain | 21 | 2M in · 0 cached · 37.8K out tokens · of them audio: 139.4K in · 20.9K out | $0.75 in · $0.75 cached · $4.5 out / 1M tokens | $2.17 | 93% | 461 ms |
| Gemini 3.8 Live Extended Thinking | Brain | 1 | 159.4K in · 0 cached · 2.1K out tokens · of them audio: 10K in · 1.3K out | $0.75 in · $0.75 cached · $4.5 out / 1M tokens | $0.161 | 7% | 178 ms |
Most expensive sessions
| When | What | Models | Est. cost |
|---|---|---|---|
| 13 Oct, 20:13 | Live call | Gemini 3.8 Live | $0.232 |
| 26 Oct, 12:12 | Live call | Gemini 3.8 Live | $0.199 |
| 19 Oct, 14:47 | Live call | Gemini 3.8 Live | $0.184 |
| 26 Oct, 17:30 | Live call | Gemini 3.8 Live Extended Thinking | $0.161 |
| 13 Oct, 16:06 | Live call | Gemini 3.8 Live | $0.149 |
| 13 Oct, 17:31 | Simulation | Gemini 3.8 Live | $0.145 |
| 13 Oct, 17:42 | Simulation | Gemini 3.8 Live | $0.143 |
| 13 Oct, 17:36 | Simulation | Gemini 3.8 Live | $0.140 |
| 26 Oct, 17:39 | Live call | Gemini 3.8 Live | $0.119 |
| 13 Oct, 17:40 | Simulation | Gemini 3.8 Live | $0.107 |
Inference limits
The Build plan's limits from LiveKit's docs · peaks on this deployment · Last 30 days
No LiveKit Inference use in this period.
How to read this
- An OpenRouter brain's cost is OpenRouter's own bill for every request, recorded since 22 September — before that, the cheapest of its providers able to take a call. Every other cost is an estimate: what was counted on each call, times the list price. Those bills are LiveKit's, Google's, MiniMax's and Alibaba Cloud's.
- A call is counted to speech-to-speech when a speech-to-speech model answered it, and to the cascade otherwise; the drawer's MiniMax previews are neither.
- Only this deployment's calls, Simulations and drawer requests are here — the laptop and staging record into different databases. Usage is recorded from the day this page was added; a worker that stops mid-call loses up to its last 20 seconds.
- LiveKit's credit and turn detector are counted this month from this deployment's calls; the other worker on the same LiveKit project spends the same credit, and LiveKit Cloud's dashboard has the real balance. The turn detector is free for an agent outside LiveKit Cloud, shared across the account's free projects.
- A Model Studio model's free quota runs 90 days from activation, not by the month, so it is counted from every call and Simulation recorded here. Calls made anywhere else — the laptop, a probe — are not in it, so it can read as half used for a quota already gone; only Model Studio's console has the real balance, and no API gives it.
- LiveKit limits are the Build plan's, from LiveKit's docs; the peaks are this deployment's own. Connections count the sessions running at once, requests and tokens the busiest 60 seconds, and the turn detector this month's requests.
- MiniMax charges a designed or cloned voice when it first speaks, not when it is made; it is counted here when it is made. Qwen's free hours in its first 90 days are not subtracted.
Where the prices and limits come from
Prices in USD at list rates, read 2026-09-11 — LiveKit Inference: livekit.com/pricing/inference, Build plan rates as of 2026-09-09 (per audio minute, per million tokens, per million characters); the Build plan's $2.50 monthly credit: livekit.com/pricing; 7,500 free turn-detector requests a month for an agent outside LiveKit Cloud: docs.livekit.io quotas and limits. Qwen: Alibaba Cloud Model Studio model pricing, international rate $0.00009 per second (a 10-hour free quota in the first 90 days is not subtracted). MiniMax: platform.minimax.io pay-as-you-go pricing — speech per million characters, a designed voice $3 and a cloned voice $1.5, each charged when the voice first speaks, a designed voice's preview at $60 per million characters; speech-01 models are not listed there. Qwen3.5-Omni Realtime: Model Studio model pricing, international (Singapore), read 2026-09-14 — per million tokens, text and audio input priced apart, a reply that speaks billed by its audio tokens alone; a free quota of 1 million tokens is not subtracted. OpenRouter brains: each provider's own price, from openrouter.ai/api/v1/models/<model>/endpoints, among the providers that take a call's tools and reasoning setting — the cheapest as the estimate, the dearest as the ceiling — read 2026-09-15. LiveKit promotions: a model on a promotion keeps its regular price, and its promotion price and end, from livekit.com/pricing/inference, read 2026-09-15. Qwen-Audio 3.0 Realtime: Model Studio model pricing, international (Singapore), read 2026-09-17 — per million tokens, plus $0.8 text and $6.4 audio in, $6.4 text and $24 audio out; flash $0.45 / $4.5 / $4.5 / $15 then, $0.23 / $0.93 / $0.7 / $1.87 when read again on 2026-10-07; a free quota of 1 million tokens is not subtracted. Model Studio free quota: each model has its own, typically 1,000,000 tokens, valid 90 days from the day Model Studio was activated — or, since read again on 2026-10-07, from the model's release where that is later — and shared by the account's RAM users — alibabacloud.com/help/en/model-studio/new-free-quota, read 2026-09-17. It cannot be read through an API: only Model Studio's console shows what is left. Gemini Live: ai.google.dev/gemini-api/docs/pricing, paid tier, read 2026-09-23 — per million tokens, text and audio priced apart; Gemini 3.8 Live and 3.1 Flash Live $0.75 text and $3 audio in, $4.5 text and $12 audio out; 3.8 Live Extended Thinking in the same row as 3.8 Live (read 2026-10-26); 2.5 Flash Live $0.5 / $3 / $2 / $12. Output includes thinking tokens, billed as text (textOutputBilled). The free tier is not counted: its terms let Google use what is sent. Qwen3.8-Omni Flash Realtime and Qwen-Audio 3.1 Realtime Plus: Model Studio model pricing, international (Singapore), read 2026-10-07 — 3.8 $0.23 text and $0.93 audio in, $0.70 text and $1.87 audio out, and it bills a spoken reply's text as well as its audio (textOutputBilled); 3.1 Plus as 3.0 Plus; a free quota of 1 million tokens each is not subtracted.
LiveKit Inference limits on the Build (free) plan, read 2026-09-12 from docs.livekit.io/deploy/admin/quotas-and-limits: 5 active STT connections and 5 active TTS connections to LiveKit Inference models; 100 LLM requests and 600,000 LLM input and output tokens in a sliding 60-second window. The turn detector's only documented limit is its 7,500 free requests a month for an agent outside LiveKit Cloud (pricing.turnDetectorFreeRequests); adaptive interruption, which these calls do not use, has 40,000.