LiveKit pricing

About

LiveKit is the open-source, end-to-end WebRTC stack that has become the default real-time transport layer for voice and video AI. Launched in July 2021 as a free, Apache-2.0 server you can self-host, it grew into a three-part product: the open-source media server, LiveKit Cloud (the managed, globally-distributed network), and the Agents framework for building voice/video AI agents. LiveKit provides the real-time transport behind OpenAI’s ChatGPT voice mode, and the Agents framework — modeled on that work — is downloaded more than a million times a month. Other customers include xAI, Salesforce (Agentforce), Tesla, and even 911 emergency services; the network handles billions of calls a year.

The company raised a $45M Series B at a $345M valuation in April 2025 (led by Altimeter), then a $100M Series C at a $1B valuation in January 2026 (led by Index Ventures, with Salesforce Ventures, Hanabi, Altimeter and Redpoint). At its Series B it reported 500+ paying customers and 100,000+ developers.

For the most current information, visit LiveKit.

Pricing summary : How LiveKit’s pricing model works

LiveKit has two pricing paths. The open-source server is free to self-host under Apache 2.0 — no LiveKit fees, you pay only for your own infrastructure. LiveKit Cloud is the managed alternative with four tiers: Build (free), Ship ($50/mo), Scale ($500/mo), and Enterprise (custom).

Cloud is multi-dimension metered. Each tier bundles monthly allotments of several units, then charges usage-based overage above them:

• Agent-session minutes — time an AI agent runs on Cloud. Build includes 1,000, Ship 5,000, Scale 50,000; overage is $0.01 per min.
• WebRTC media minutes — end-user connection time to the realtime network. Build 5,000, Ship 150,000, Scale 1.5M; overage is $0.0005 per min (Ship) / $0.0004 per min (Scale).
• Inference credits — for LiveKit Inference (LLM/STT/TTS via one API key). Build includes $2.50 in credits (~50 minutes), Ship $5 (~100 minutes), Scale $50 (~1,000 minutes, then billed at discounted model prices).
• Agent observability — session recordings (1,000 / 5,000 / 50,000 min included, then $0.005 per min) and observability events (100,000 / 500,000 / 5,000,000 entries, then $0.00003 per entry).
• Telephony — 1 free US local number, then $1.00/month per number and $0.01 per inbound min; toll-free is $2.00/month per number and $0.02 per minute; third-party SIP minutes run $0.004 per min (Ship) / $0.003 per min (Scale).
• Data transfer — 50 GB / 250 GB / 3 TB included, then $0.12 per GB (Ship) / $0.10 per GB (Scale).

Two smaller meters ride alongside: voice isolation minutes (100 / 1,000 / 10,000 included, then $0.0012/min) and hard concurrency ceilings — concurrent agent sessions (5 / 20 / up to 600, starting at 50 with more on request), inference concurrency (5 / 20 / 50) and concurrent connections (100 / 1,000 / 5,000).

What makes this different: LiveKit prices the AI-agent workload, not just raw video conferencing — a textbook hybrid pricing model with a small subscription floor and a wide metered surface. The headline meter — agent-session minutes at a flat $0.01 per min — is independent of which LLM/STT/TTS models you call (those bill separately through inference credits or your own provider keys). A built-in calculator estimates blended per-minute agent cost — its default Gemma 4 31B + Deepgram Nova-3 (Multilingual) + Cartesia Sonic 3 voice stack with observability enabled lands at $0.0672/min — so buyers can model a voice agent end-to-end before committing.

Pricing by product

LiveKit Cloud (plan tiers)

LiveKit Cloud (metered units and overage)

Enterprise is “Custom” on every row above. Export of recordings, transcripts, traces and logs to cloud storage is listed as Coming soon on Ship, Scale and Enterprise (and unavailable on Build), and custom SIP domains are not offered on any published tier.

LiveKit Inference (model prices, per minute)

Inference draws down the plan’s credits, then bills per minute of model time. Scale (and Enterprise) get discounted STT/TTS rates; LLM rates are published as a single per-minute list.

Fish Audio is a new TTS vendor added to the catalog, including a free SKU (S2.1 Pro Free, listed as Free) with no Scale discount. Two new LLM entries also joined the list, Google Gemini 3.5 Flash Lite ($0.0013/min) and Gemini 3.6 Flash ($0.0058/min). 2026-08-13: Deepgram adds a second TTS SKU, Flux TTS, priced free ($0.00/min) on both Build/Ship and Scale — the catalog’s third $0/min model alongside Fish Audio’s S2.1 Pro Free; two duplicate LLM SKUs, GPT-5.2 Chat and GPT-5.3 Chat (each priced like their non-Chat base model, ~$0.0077/min), are removed from the published LLM list; and Gemini 3.6 Flash’s cached-input token rate is disclosed for the first time at $0.150 per million tokens (previously blank/N/A). 2026-08-14: one new LLM SKU joins the catalog, Google Gemini 3.7 Flash, at $0.0029/min ($0.750 input / $0.075 cached input / $3.750 output per million tokens) — priced between Gemini 3.6 Flash ($0.0058/min) and Gemini 3.5 Flash Lite ($0.0013/min) on the published list; no Cloud plan price, allotment or overage rate moved. 2026-08-26: xAI’s Grok 4.1 Fast and Grok 4.1 Fast Reasoning (each formerly ~$0.0007/min, $0.200 input / $0.500 output per million tokens) are removed from the published LLM list, trimming xAI’s lineup to five models (Grok 4.20, Grok 4.20 Reasoning, Grok 4.20 Multi-Agent, Grok 4.3, Grok 4.5); on the TTS side, Rime’s Arcana voice (formerly ~$0.0240/min Build-Ship, ~$0.0180/min Scale) is delisted and a new provider, Gradium, joins with a single Gradium TTS voice priced higher at $0.0288/min (Build/Ship) and $0.0216/min (Scale); no Cloud plan price, allotment or overage rate moved. 2026-08-28: two new STT SKUs join the catalog — Google Gemini 3.5 Transcribe Live at $0.0095/min and Speechmatics Linden-1 at $0.0050/min, both priced identically on Build/Ship and Scale (no Scale discount) — and the vendor label for xAI’s models (the Grok LLM family plus its Speech to Text and Text to Speech entries) is relabeled SpaceXAI throughout the LLM, STT and TTS tables, with no per-model rate change for the renamed vendor; no Cloud plan price, allotment or overage rate moved. 2026-09-07: ElevenLabs is delisted entirely from the catalog — its one STT SKU (Scribe v2 Realtime) and all six TTS voices (Eleven Flash v2, Flash v2.5, Multilingual v2, Turbo v2, Turbo v2.5, Eleven v3) no longer appear on either the main pricing page or the dedicated Inference page; Rime’s four voices (Coda, Mist, Mist v2, Mist v3), formerly $50.00/$30.00/$30.00/$30.00 per million characters on Build/Ship and $50.00/$20.00/$20.00/$20.00 on Scale, are now Free ($0.00) on every tier; two new LLM SKUs join (Google Gemini 3.8 Flash $0.0029/min, SpaceXAI Grok 4.6 $0.0070/min) alongside two new TTS SKUs (Cartesia Sonic 3.6, Inworld Realtime TTS 2.0 Flash); a duplicate AssemblyAI STT SKU (Universal-3 Pro Streaming) is removed; and LiveKit Cloud’s metered-units table gains a new Non-production deployments row (Build 0, Ship 2, Scale 5, Enterprise Custom); no Cloud plan price, allotment or overage rate moved.

Sales motions across products: self-serve PLG for Build/Ship/Scale (instant signup, no card on Build), open-source self-host, and sales-led for Enterprise (volume + inference discounts, on-prem/private deployment).

Hidden costs : What LiveKit users actually pay

The flat $50/$500 is only the floor. A production voice-AI app stacks five separate meters: agent-session minutes, WebRTC media minutes, inference credits (or your own model bills), telephony, and data transfer. The single biggest line is usually inference — the LLM/STT/TTS model minutes that the LiveKit calculator surfaces — which can dwarf the $0.01/min agent-session fee. For example, the calculator’s default voice stack (Gemma 4 31B + Deepgram Nova-3 Multilingual + Cartesia Sonic 3 + observability) lands at $0.0672/min all-in, of which the LiveKit agent-session + observability portion is only $0.02/min — swap the LLM for OpenAI GPT Realtime and the same session costs roughly twice as much.

Other things to budget for: HIPAA, RBAC, region pinning and metrics-export APIs are gated to Scale ($500) and above; cold-start prevention (always-on agents), custom voices and inference discounts also start at Ship/Scale; and toll-free numbers ($2/number, $0.02/inbound min) and third-party SIP minutes bill on top.

Want to estimate your own LiveKit bill? Use the LiveKit pricing calculator to model your costs based on usage patterns.

Pricing evolution : LiveKit pricing history and changes

Cadence

Tracked range: 2021–2026 Q3. Periods not listed above carried no plan-price changes and no SKU additions.

Notable changes

• 2021-07 — Launches as a free, open-source, end-to-end WebRTC stack (Apache 2.0). Self-hosting carries no LiveKit fees.
• 2023-09 — LiveKit powers OpenAI’s ChatGPT voice mode and releases the open-source Agents framework; LiveKit Cloud (Build/Ship/Scale) matures, metered on participant/connection minutes and bandwidth.
• 2025-04 — $45M Series B at $345M (FinSMEs). Cloud repositions around voice/video AI agents, making agent-session minutes and inference credits the primary metered units.
• 2026-01 — $100M Series C at a $1B valuation led by Index Ventures (LiveKit blog; TechCrunch).
• 2026-06 — Current structure: Build (free), Ship $50, Scale $500, Enterprise. Seven metered dimensions with per-unit overage — agent sessions, WebRTC minutes, inference credits, observability (recordings and events), voice isolation, telephony and data transfer — with LiveKit Inference exposing LLM/STT/TTS model minutes through one API key.
• 2026-07-21 — Six new LLM SKUs land in LiveKit Inference (GPT-5.6 Luna $0.0040/min, Terra $0.0101/min, Sol $0.0203/min; Grok 4.3 $0.0042/min, Grok 4.5 $0.0070/min; Kimi K2.6 $0.0035/min) with no change to any plan price, allotment or overage rate. The visible move is on the pricing-page calculator, whose default LLM switched from GPT-5.3 Chat ($0.0077/min) to Gemma 4 31B ($0.0014/min), cutting the advertised blended voice-agent estimate from $0.0735/min to $0.0672/min — a 9% drop in the headline number that LiveKit achieved by changing a default rather than a price.
• 2026-07-29 — Fish Audio joins LiveKit Inference as a new text-to-speech vendor, eight days after the previous catalog update, with three SKUs — S2 Pro ($0.0090/min), S2.1 Pro ($0.0090/min) and a free S2.1 Pro Free ($0.0000/min) — none discounted on Scale, unlike most other TTS entries on the list. Two Gemini LLM entries also join, Google Gemini 3.5 Flash Lite ($0.0013/min) and Gemini 3.6 Flash ($0.0058/min). No Cloud plan price, allotment or overage rate moved.
• 2026-08-11 — Two existing LiveKit Inference LLM SKUs get cheaper: GPT-5.6 Luna falls from $0.0040/min to $0.0008/min (an 80% cut) and GPT-5.6 Terra falls from $0.0101/min to $0.0081/min (a 20% cut); GPT-5.6 Sol holds at $0.0203/min. This is a genuine per-SKU repricing, not a calculator-default swap — the pricing-page calculator’s advertised blended estimate still reads $0.0672/min on its default Gemma 4 31B stack, unaffected because Gemma wasn’t one of the repriced models. No Cloud plan price, allotment or overage rate moved.
• 2026-08-14 — One new LLM SKU joins LiveKit Inference: Google Gemini 3.7 Flash at $0.0029/min ($0.750 input / $0.075 cached input / $3.750 output per million tokens), slotting between Gemini 3.6 Flash ($0.0058/min) and Gemini 3.5 Flash Lite ($0.0013/min) on the published per-minute list. No Cloud plan price, allotment or overage rate moved.
• 2026-08-26 — LiveKit Inference reshuffles two corners of the catalog: xAI’s Grok 4.1 Fast and Grok 4.1 Fast Reasoning (each formerly ~$0.0007/min) are delisted, trimming xAI’s LLM lineup to five models; Rime’s Arcana TTS voice is dropped and a new provider, Gradium, joins with a single Gradium TTS voice at $0.0288/min (Build/Ship) / $0.0216/min (Scale). No Cloud plan price, allotment or overage rate moved.
• 2026-08-28 — Two new STT SKUs join LiveKit Inference and xAI is relabeled SpaceXAI: Google Gemini 3.5 Transcribe Live ($0.0095/min) and Speechmatics Linden-1 ($0.0050/min) join the catalog, both priced identically on Build/Ship and Scale. Separately, the catalog’s vendor label for xAI’s models — the Grok LLM family plus its Speech to Text and Text to Speech entries — changes from “xAI” to “SpaceXAI” across the LLM, STT and TTS tables, with no per-model rate change. No Cloud plan price, allotment or overage rate moved.
• 2026-09-07 — LiveKit delists ElevenLabs entirely from LiveKit Inference and cuts Rime’s TTS voices to $0.00/min. ElevenLabs’ one STT SKU (Scribe v2 Realtime) and six TTS voices are removed from both pricing pages with no stated migration path; Rime’s four voices (Coda, Mist, Mist v2, Mist v3) drop from $20–$50 per million characters to Free on every tier — the catalog’s largest tier-wide price cut to date. Two new LLM SKUs join (Google Gemini 3.8 Flash $0.0029/min, SpaceXAI Grok 4.6 $0.0070/min) alongside two new TTS SKUs (Cartesia Sonic 3.6, Inworld Realtime TTS 2.0 Flash), and a duplicate AssemblyAI STT SKU is cleaned up. LiveKit Cloud’s metered-units table also gains a new “Non-production deployments” allotment row (Build 0 / Ship 2 / Scale 5 / Enterprise Custom) — a per-tier cap with no published overage rate, so the billable meter count stays at seven. No Cloud plan price, allotment or overage rate moved.

The ElevenLabs delisting in detail

ElevenLabs had been in LiveKit’s Inference catalog across the whole tracked history of the current pricing structure — listed as far back as 2026-06-09 with six TTS voices (Eleven Flash v2, Flash v2.5, Multilingual v2, Turbo v2, Turbo v2.5, Eleven v3) and one STT model (Scribe v2 Realtime, $0.0105/min), and still listed on 2026-08-28. On 2026-09-07 all seven SKUs disappeared from both livekit.io/pricing and the dedicated livekit.com/pricing/inference page, in the same release that added Google Gemini 3.8 Flash, SpaceXAI Grok 4.6, Cartesia Sonic 3.6 and Inworld Realtime TTS 2.0 Flash. Nothing in the pricing-page copy names a deprecation window, a suggested-replacement voice, or a credit-based makegood for customers whose agents were pinned to a specific ElevenLabs voice — the removal reads identically to every other routine catalog refresh LiveKit has shipped since July. That is the risk of a pass-through inference meter: it is exactly as easy for LiveKit to drop a vendor as to add one, and unlike a price change (which shows up as a bigger bill) a delisting shows up as a broken integration.

What’s unique : LiveKit’s distinctive pricing mechanics

1. Agent-session minutes as the headline meter. Rather than billing on raw video minutes or seats, LiveKit prices the agent runtime at a flat $0.01/min — decoupled from which models you run. It’s a value metric that maps directly to “how long my voice agent was live,” which is intuitive for AI builders.

2. Genuinely free via open source. The full WebRTC server is Apache 2.0 and self-hostable with no LiveKit fees. Cloud sells the managed global network, agent deployment/observability, and inference convenience — not the core capability — which caps pricing power but maximizes adoption.

3. Model-cost passthrough with optional discounts. LiveKit Inference lets you call LLM/STT/TTS models through one key, billed via credits; Scale and Enterprise get discounted model rates. You can also bring your own provider keys. This turns model spend into a metered, optionally-marked-up dimension layered on the subscription — and the catalog is restocked continuously rather than repriced: the 2026-07-21 refresh added six LLM SKUs (GPT-5.6 Luna/Terra/Sol, Grok 4.3/4.5, Kimi K2.6) while leaving every LiveKit-owned rate untouched. The published LLM list now spans roughly $0.0002/min to $0.0676/min, so which model you pick moves the bill by two orders of magnitude while LiveKit’s own take stays a flat $0.01/min. Eight days later, on 2026-07-29, the same pattern repeated on the TTS side: Fish Audio joined as a brand-new vendor with three SKUs, including a free S2.1 Pro Free tier — the catalog’s first $0/min model — while Gemini 3.5 Flash Lite and Gemini 3.6 Flash extended the low end of the LLM list. Notably, none of the three Fish Audio SKUs carry the Scale discount that most other TTS vendors get, which shows the “discounted model rates” promise is applied vendor-by-vendor rather than as a blanket Scale benefit. The catalog kept moving through August without breaking that pattern: a same-SKU GPT-5.6 repricing on 2026-08-11, a second free TTS SKU (Deepgram Flux TTS) on 2026-08-13, and a fifth LLM entry, Google Gemini 3.7 Flash at $0.0029/min, on 2026-08-14 — none of which touched a Cloud plan price, allotment, or overage rate. The pattern held into late August too: on 2026-08-26, xAI’s Grok 4.1 Fast and Fast Reasoning SKUs were delisted and Rime’s Arcana TTS voice was swapped for a new provider, Gradium; on 2026-08-28, two new STT SKUs joined — Google Gemini 3.5 Transcribe Live and Speechmatics Linden-1, both priced flat across Build/Ship and Scale with no Scale discount, extending the non-discounted-exception pattern first seen with Fish Audio — while the catalog’s xAI listings were simultaneously relabeled “SpaceXAI” with no rate change for the renamed vendor. On 2026-09-07, the pattern crossed a new threshold: rather than trimming a couple of SKUs or swapping one small vendor for another, LiveKit delisted an entire established vendor — ElevenLabs, with one STT and six TTS SKUs — from the catalog outright, while simultaneously cutting Rime’s four TTS voices to $0.00/min on every tier (the catalog’s largest tier-wide price cut to date) and adding two more LLM SKUs (Gemini 3.8 Flash, Grok 4.6) and two more TTS SKUs (Cartesia Sonic 3.6, Inworld Realtime TTS 2.0 Flash). Eight catalog updates in about seven weeks, still zero Cloud plan repricing — but the first time a multi-SKU vendor has been removed wholesale rather than incrementally adjusted.

4. The calculator default is itself a pricing lever. LiveKit advertises a blended per-minute number rather than a plan price, and that number is a function of the models pre-selected in the on-page estimator. On 2026-07-21 the default LLM moved from GPT-5.3 Chat to Gemma 4 31B and the advertised total fell from $0.0735/min to $0.0672/min without a single rate changing. It is an honest number — the stack is named on screen — but it means the headline moves with merchandising decisions, and two quotes taken months apart are not comparable unless you fix the model stack first.

Strengths & weaknesses

Billing UX : LiveKit billing controls and transparency

• Self-serve signup, “No credit card required” — Build starts free with 1,000 free agent-session minutes monthly and no card; upgrades to Ship ($50/mo) and Scale ($500/mo) are self-serve from the same flow.
• Pricing calculator (“Estimate costs for AI voice and video agents”) — an interactive per-minute estimator on the pricing page with a How users connect toggle (Phone call / Web-mobile) and a Select a plan toggle (Build/Ship vs Scale), breaking the bill into Agent session, Telephony, WebRTC connection, LLM, STT, TTS and Observability lines and summing a Total estimated cost ($0.0672/min on the default stack).
• “View pricing in plain text (Markdown)” — a machine-readable dump of the whole pricing page, linked from the top of the pricing page for agents and scripts.
• Per-second metering, 10-second minimum — since August 2026, agent-session minutes and recordings, WebRTC connections, and SIP connections are billed per second (not rounded up to the full minute), with a 10-second minimum per session; LiveKit says this applies to every plan, including Enterprise contracts, and first showed up on September 2026 invoices for August usage. None of the published per-unit rates changed — a short outbound call that reaches an answering machine now costs a few cents instead of a full minute’s worth.
• Concurrency ceilings with a dashboard raise path — concurrent agent sessions, LiveKit Inference concurrency and concurrent connections are each hard-capped per tier, with “request more via dashboard” as the documented lift for Scale.
• Agent observability + deployment metrics — agent session recordings, turn-by-turn observability events (transcripts, trace spans, logs), deployment metrics (resource allocation, latency, errors) and session metrics/analytics in-product; metrics export APIs unlock at Scale, and export to cloud storage is marked Coming soon.
• Zero data retention — available on every published tier: prompts, audio and model outputs aren’t logged or stored by LiveKit or the underlying model providers.
• Payment options — card-based self-serve for Build/Ship/Scale; Enterprise is quoted by sales (“Contact sales”) with volume pricing including inference, SSO, a shared Slack channel and a support SLA.

Strategic wins : Why LiveKit’s pricing decisions worked

1. Open source as the top-of-funnel

Launching a free Apache-2.0 WebRTC stack made LiveKit the default real-time layer for an entire generation of voice/video apps — including OpenAI’s ChatGPT voice mode. The free self-host path removes adoption friction; Cloud monetizes the teams that don’t want to operate a global media network. See how AI companies structure pricing.

2. Repricing around the agent, not the call

By moving the headline meter to agent-session minutes, LiveKit aligned its pricing with the unit AI builders actually reason about, riding the voice-AI wave from a $345M (April 2025) to a $1B (January 2026) valuation. Related: outcome-based pricing trends.

3. Inference as a layered, discountable meter

Folding LLM/STT/TTS access into metered inference credits (with Scale/Enterprise discounts) lets LiveKit capture model spend as an expansion lever without forcing a single model choice. Because the meter is the minute, not the model, LiveKit can restock the catalog as fast as the labs ship — the 2026-07-21 update dropped in GPT-5.6, Grok 4.3/4.5 and Kimi K2.6 with zero repricing work and zero migration for existing customers. That turns “we support the newest model” into routine catalog maintenance instead of a pricing event. The 2026-07-29 update repeated the pattern on the TTS side eight days later — onboarding an entirely new vendor, Fish Audio, complete with a free SKU, plus two Gemini LLM entries — underscoring that catalog expansion, not plan repricing, is now LiveKit’s default move between major releases. Six more updates followed the same script through early September — a targeted GPT-5.6 repricing (08-11), a free Deepgram Flux TTS SKU (08-13), a new LLM entry, Google Gemini 3.7 Flash (08-14), a Grok 4.1 Fast delisting paired with a Rime-to-Gradium TTS swap (08-26), two new STT SKUs alongside an xAI-to-SpaceXAI vendor relabel (08-28), and on 2026-09-07 the wholesale delisting of ElevenLabs (one STT plus six TTS SKUs) alongside Rime’s four TTS voices cutting to $0.00/min and four new SKUs (Gemini 3.8 Flash, Grok 4.6, Cartesia Sonic 3.6, Inworld Realtime TTS 2.0 Flash) — eight catalog moves in about seven weeks with zero Cloud plan repricing across any of them. The lever cuts both ways: the same restock cadence that adds a model for free can drop one for free too, so a strategic win on flexibility for LiveKit is a standing integration risk for any customer pinned to a specific vendor. See choosing the right usage metric.

Areas to improve : Gaps in LiveKit’s pricing approach

1. Too many meters to forecast confidently

Seven simultaneous metered dimensions (agent minutes, WebRTC minutes, inference, observability recordings, observability events, voice isolation, telephony and bandwidth) make it genuinely hard to predict a monthly bill without running the calculator for each scenario — and observability quietly doubles LiveKit’s own per-minute take, since the calculator’s $0.01/min observability line matches the agent-session line. A blended “all-in per-minute” headline, or a saved-scenario feature in the calculator, would reduce planning friction. See bill shock and cost unpredictability.

2. Compliance gated high

HIPAA, RBAC, region pinning and metrics-export APIs only appear at the $500 Scale tier, which can push regulated startups straight to a steep step-up well before their volume justifies it.

3. Model-cost dependence — and a moving headline number

Because inference is usually the dominant line and rides third-party model pricing, LiveKit’s effective cost-to-serve for a customer can swing with provider price changes — a transparency and predictability gap LiveKit only partly controls. The 2026-07-21 refresh showed the buyer-facing edge of that: the advertised blended estimate dropped from $0.0735/min to $0.0672/min because the calculator’s default LLM changed, not because anything got cheaper for an existing customer still on GPT-5.3 Chat. A durable fix is cheap — stamp the calculator with the model stack and date it assumes, and let buyers save or share a fixed configuration — so a quote made in June is still comparable to one made in September.

4. Scale’s “discounted model prices” promise isn’t universal

The pricing page markets Scale’s larger inference credit line as unlocking discounted model prices, and most STT/TTS vendors do get a materially lower Scale rate (Deepgram Nova-3 $0.0058→$0.0050/min, Cartesia Sonic 3 $0.0300→$0.0225/min). Fish Audio, added 2026-07-29, breaks that pattern — all three of its SKUs price identically on Ship and Scale. Two new STT SKUs added 2026-08-28 — Google Gemini 3.5 Transcribe Live and Speechmatics Linden-1 — repeat the same pattern, pricing flat on Build/Ship and Scale; the exception list is now five SKUs across two separate catalog updates, not a one-off. A buyer upgrading to the $500/mo Scale plan specifically for the inference discount should not assume it applies to every model; a per-vendor “Scale discount” flag on the pricing-page catalog would prevent that assumption from costing $450/mo more than expected for no benefit.

5. Vendor delisting risk has no visible migration path

On 2026-09-07, LiveKit dropped ElevenLabs from the Inference catalog entirely — one STT SKU and six TTS voices, gone from both the main and dedicated Inference pricing pages with no announced grace period, discount, or suggested-replacement voice. Any customer who built a voice agent around a specific ElevenLabs voice now has to re-record, re-test, and re-tune with a different provider on their own timeline, not LiveKit’s. Because LiveKit Inference’s whole pitch is “call the model, we handle the plumbing,” a delisting like this is the mirror image of its greatest strength — the same mechanism that lets LiveKit add a vendor without a pricing event lets it remove one the same way. A durable fix is to publish a deprecation window (e.g., 30–60 days) and a suggested-replacement mapping whenever a vendor is dropped, the way cloud providers announce API sunsets, so buyer impact shows up as advance notice rather than a broken integration.

Key takeaways

1. Open source built the funnel. A free Apache-2.0 WebRTC stack made LiveKit the default real-time layer for voice AI — including ChatGPT voice mode — before Cloud monetized the managed network.
2. Price the agent, not the call. Switching the headline meter to agent-session minutes ($0.01/min) aligned pricing with the AI-builder’s mental model and tracked a 3x valuation jump in nine months.
3. A flat floor plus many meters trades simplicity for fairness. Build/Ship/Scale anchor the bill, but seven overage dimensions make forecasting hard.
4. Inference is the real cost driver, and defaults — and vendor lineups — are prices. The $0.01/min agent fee is small next to blended model minutes (~$0.04–$0.07/min) on a published LLM list running $0.0002–$0.0676/min. LiveKit cut its advertised blended estimate 9% on 2026-07-21 by changing the calculator’s default LLM rather than any rate, and on 2026-09-07 it deleted an entire vendor (ElevenLabs, 7 SKUs) from the catalog outright — whatever a buyer sees pre-selected, and whichever vendors remain listed, are effectively your list price and your product surface.
5. Self-host stays the escape valve. Because the core server is free under Apache 2.0, Cloud has to win on convenience and scale, not lock-in.

UBP implications

1. Choose a value metric your buyer already counts. “Agent-session minutes” maps cleanly to how voice-AI teams think about runtime — a more intuitive meter than raw video minutes or seats.
2. Meter the minute, not the model — then the catalog is free to churn. Passing through (and discounting at scale) third-party model spend lets a platform expand revenue without dictating a model choice, and because the unit is time rather than a named SKU, adding GPT-5.6, Grok 4.5 and Kimi K2.6 on 2026-07-21 required no repricing, no plan change and no customer migration — a pattern LiveKit repeated eight days later by onboarding an entirely new TTS vendor, Fish Audio, on 2026-07-29 the same way, then again with a GPT-5.6 repricing (08-11), a free TTS SKU (08-13), a new LLM entry, Google Gemini 3.7 Flash (08-14), a Grok 4.1 Fast delisting and Rime-to-Gradium TTS swap (08-26), and two new STT SKUs (08-28). Even a vendor rebrand rode the same channel — xAI’s catalog listings were relabeled “SpaceXAI” on 2026-08-28 with no rate change anywhere. The 2026-09-07 update went further than a routine SKU swap: it deleted an entire vendor, ElevenLabs (one STT plus six TTS voices), from the catalog outright — the first time a multi-SKU vendor was removed wholesale rather than one model being swapped for another — while simultaneously cutting Rime’s four TTS voices to $0.00/min, LiveKit’s largest tier-wide markdown yet. Eight catalog moves in about seven weeks with zero Cloud plan repricing is stronger evidence that the meter design absorbs not just new supply but also vendor exits — a UBP reseller passing through fast-moving upstream supply should plan for full vendor churn, not just price and SKU churn, and should publish a migration path for buyers when a vendor’s models disappear entirely.
3. Open core changes pricing power. When the engine is free to self-host, the paid tiers must sell the managed network, observability, and compliance — not the capability. See usage-based pricing strategy.

Sources

• LiveKit pricing page (accessed 2026-09-07)
• LiveKit GitHub (open-source WebRTC server, Apache 2.0) (accessed 2026-06-09)
• LiveKit Series C: towards the voice-driven era of computing (accessed 2026-06-09)
• LiveKit Inference pricing (per-model LLM, STT and TTS rates) (accessed 2026-09-07)
• LiveKit documentation — quotas and limits (accessed 2026-08-28)
• LiveKit documentation — LLM models overview (accessed 2026-08-28)
• LiveKit blog — Introducing per-second metering for LiveKit (accessed 2026-09-08)

Bottom line

LiveKit is the open-source (Apache 2.0) WebRTC stack that became the default real-time layer for voice and video AI — it powers OpenAI’s ChatGPT voice mode and its Agents framework is downloaded over a million times a month. The server is free to self-host; LiveKit Cloud is a hybrid — free Build, $50 Ship, $500 Scale and custom Enterprise — metered across seven dimensions including agent-session minutes ($0.01/min), WebRTC media minutes ($0.0004–$0.0005/min), inference credits, observability, telephony and data transfer. Plan rates have held steady through 2026; the movement is in the LiveKit Inference catalog, which added GPT-5.6, Grok 4.3/4.5 and Kimi K2.6 on 2026-07-21, a new TTS vendor (Fish Audio, including a free SKU) and two Gemini LLM entries on 2026-07-29, a GPT-5.6 repricing on 2026-08-11, a free Deepgram Flux TTS SKU on 2026-08-13, a new LLM entry, Google Gemini 3.7 Flash, on 2026-08-14, a Grok 4.1 Fast delisting paired with a new Gradium TTS provider on 2026-08-26, two new speech-to-text SKUs alongside an xAI-to-SpaceXAI vendor rename on 2026-08-28, and on 2026-09-07 the wholesale delisting of ElevenLabs (7 SKUs, no stated migration path) alongside Rime’s four TTS voices cut to $0.00/min and four new SKUs (Gemini 3.8 Flash, Grok 4.6, Cartesia Sonic 3.6, Inworld Realtime TTS 2.0 Flash) — eight catalog updates in about seven weeks with no Cloud plan repricing — while the pricing-page calculator’s advertised blended voice-agent estimate holds at $0.0672/min. After a $45M Series B at $345M (April 2025), LiveKit raised a $100M Series C at a $1B valuation in January 2026. Browse the pricing blueprint for more fully-researched company profiles.

Want to compare LiveKit against other voice and real-time AI companies? Browse the pricing blueprint.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

ElevenLabs delisted; Rime TTS goes free in LiveKit Inference

Sep 2026

Plan structure holds (Build free / Ship $50 / Scale $500 / Enterprise custom). LiveKit Inference delists ElevenLabs entirely (1 STT + 6 TTS SKUs) and cuts Rime's four TTS voices to $0.00/min on every tier; adds Google Gemini 3.8 Flash and SpaceXAI Grok 4.6 (LLM) plus Cartesia Sonic 3.6 and Inworld Realtime TTS 2.0 Flash (TTS); removes a duplicate AssemblyAI STT SKU; and Cloud's metered-units table gains a new 'Non-production deployments' allotment row (Build 0 / Ship 2 / Scale 5 / Enterprise Custom), a per-tier cap with no overage rate.

Two new STT SKUs join LiveKit Inference; xAI relabeled SpaceXAI

Aug 2026

Plan structure holds (Build free / Ship $50 / Scale $500 / Enterprise custom). LiveKit Inference gains two speech-to-text SKUs — Google Gemini 3.5 Transcribe Live ($0.0095/min) and Speechmatics Linden-1 ($0.0050/min), both flat across Build/Ship and Scale with no Scale discount — while the catalog's xAI vendor label (Grok LLM family plus its STT/TTS rows) is renamed SpaceXAI throughout, with no per-model rate change.

Grok 4.1 Fast delisted; Rime Arcana swapped for Gradium TTS

Aug 2026

Plan structure holds (Build free / Ship $50 / Scale $500 / Enterprise custom). LiveKit Inference delists xAI's Grok 4.1 Fast and Grok 4.1 Fast Reasoning LLM SKUs (each formerly ~$0.0007/min), trimming xAI's published lineup to five models, and swaps Rime's Arcana TTS voice for a new provider, Gradium, at $0.0288/min (Build/Ship) / $0.0216/min (Scale).

Google Gemini 3.7 Flash added to LiveKit Inference

Aug 2026

Plan structure holds (Build free / Ship $50 / Scale $500 / Enterprise custom). LiveKit Inference gains one new LLM SKU, Google Gemini 3.7 Flash, at $0.0029/min ($0.750 input / $0.075 cached input / $3.750 output per million tokens), slotting between Gemini 3.6 Flash and Gemini 3.5 Flash Lite on the published list.

Deepgram Flux TTS added free; GPT-5.2/5.3 Chat removed from LiveKit Inference

Aug 2026

Plan structure holds (Build free / Ship $50 / Scale $500 / Enterprise custom). LiveKit Inference gains a new free TTS SKU, Deepgram Flux TTS ($0.00/min on Build/Ship and Scale), and drops two duplicate LLM SKUs, GPT-5.2 Chat and GPT-5.3 Chat (each formerly ~$0.0077/min). Gemini 3.6 Flash's cached-input token rate is newly disclosed at $0.150/M tokens.

GPT-5.6 Luna and Terra repriced down in LiveKit Inference

Aug 2026

Plan structure holds (Build free / Ship $50 / Scale $500 / Enterprise custom). Two existing LiveKit Inference LLM SKUs get cheaper: GPT-5.6 Luna falls from $0.0040/min to $0.0008/min (-80%) and GPT-5.6 Terra falls from $0.0101/min to $0.0081/min (-20%); GPT-5.6 Sol holds at $0.0203/min.

Fish Audio TTS and two Gemini LLM entries join LiveKit Inference

Jul 2026

Plan structure holds (Build free / Ship $50 / Scale $500 / Enterprise custom). LiveKit Inference gains a new TTS vendor, Fish Audio — S2 Pro $0.0090/min, S2.1 Pro $0.0090/min, and a free S2.1 Pro Free at $0.0000/min, none discounted on Scale — plus two Google Gemini LLM entries, Gemini 3.5 Flash Lite $0.0013/min and Gemini 3.6 Flash $0.0058/min.

LiveKit Inference adds GPT-5.6, Grok 4.3/4.5 and Kimi K2.6

Jul 2026

Plan structure and every plan rate hold (Build free / Ship $50 / Scale $500 / Enterprise custom). The metered LiveKit Inference catalog gains six new per-minute LLM SKUs — OpenAI GPT-5.6 Luna $0.0040/min, Sol $0.0203/min, Terra $0.0101/min, xAI Grok 4.3 $0.0042/min, Grok 4.5 $0.0070/min and Moonshot Kimi K2.6 $0.0035/min — and the pricing-page calculator now defaults to a cheaper Gemma 4 31B stack at $0.0672/min.

Build free / Ship $50 / Scale $500 + multi-dimension metering

Jun 2026

Current structure: Build (free), Ship $50/mo, Scale $500/mo, Enterprise custom. Each bundles agent-session minutes, WebRTC media minutes, inference credits, telephony minutes and data transfer, then usage-based overage (agent sessions $0.01/min; WebRTC $0.0004–$0.0005/min; transfer $0.10–$0.12/GB).

Series B + voice-AI agent repositioning

Apr 2025

Raised $45M Series B at a $345M valuation (Altimeter). Cloud repositions around voice/video AI agents, foregrounding agent-session minutes and inference credits as primary metered units.

LiveKit Cloud + ChatGPT voice mode

Sep 2023

LiveKit Cloud (managed) matures with Build/Ship/Scale tiers metered on participant/connection minutes and bandwidth. LiveKit powers OpenAI's ChatGPT voice mode and releases the open-source Agents framework.

Open-source WebRTC stack launches

Jul 2021

LiveKit launches as a free, open-source, end-to-end WebRTC stack (Apache 2.0) for real-time audio/video — self-hostable with no LiveKit fees.