LiveKit vs Agora Pricing: Complete 2026 Cost Analysis

Why Fora Soft wrote this playbook

We’ve shipped real-time video and AI-agent products on both LiveKit and Agora for production clients serving millions of monthly sessions. Perspire, a live-fitness platform that raised $1.4M alongside an Oxford partnership, runs group classes and 1-on-1 sessions on a WebRTC stack we built and tuned. BrainCert, a virtual-classroom LMS, serves 1M+ learners with interactive whiteboards on the same class of infrastructure; it pulls $10M in annual revenue and has won the Brandon Hall Gold four times.

We also wrote the longer LiveKit AI Agent Development guide and the multimodal AI agents with LiveKit deep-dive, plus the Agora.io alternative review. This piece is a focused LiveKit vs Agora pricing comparison: what you pay, how you pay it, and where the invoice blows up.

Because we use Agent Engineering internally — AI-assisted pair programming across the stack — our scoping and migration numbers below reflect 2026 shipping velocity, not 2022. That’s why our estimates in the mini-case below are 30–40% tighter than typical agency quotes.

The short answer — LiveKit vs Agora pricing at a glance

For AI-agent workloads — voice agents, live interpreters, ambient meeting bots — the LiveKit vs Agora pricing gap boils down to this: Agora Conversational AI Engine bills around $0.0265 per participant-minute (all-in, audio pipeline), and LiveKit Cloud bills around $0.01 per agent session minute plus participant minutes at $0.0004 plus model inference you pass through. Plain RTC minutes (no AI) are cheaper on Agora ($0.99 per 1,000 audio minutes) than LiveKit’s participant minutes at list (~$0.40–$0.50 per 1,000 minutes).

Translation: if you’re building pure peer-to-peer video calling, Agora’s RTC quote looks tighter. If you’re building anything that runs a turn-taking AI agent on the server side — which is why most teams arrive at this comparison in 2026 — LiveKit usually lands cheaper after inference cost is included honestly.

The 2026 verdict in one paragraph

Here’s the one-line LiveKit vs Agora pricing verdict. If you’re building an AI voice or video agent in 2026, default to LiveKit Cloud — cheaper per minute, transparent pricing, HIPAA BAA on Scale, pluggable LLM. If your app is pure RTC at global scale and most of your traffic originates in APAC or LATAM, Agora still has edge-presence advantages worth paying for. If you’re consistently running above 500 concurrent sessions and already have Kubernetes expertise on staff, self-hosted LiveKit quietly becomes the most profitable path. Everything else in this article is the math that supports this one paragraph.

How each platform actually charges you

Before you compare dollar figures, understand the billing unit. LiveKit and Agora both meter participant-minutes, but everything that sits on top — AI agents, recording, telephony, inference — uses a different unit. Getting the unit wrong by 2x is a common source of sticker shock on invoice #2.

LiveKit Cloud: quotas plus linear overages

LiveKit Cloud sells you a monthly plan (Build, Ship, Scale, or Enterprise) that bundles a quota of agent-session minutes, WebRTC participant minutes, bandwidth egress, and LiveKit Inference credits. Anything above the quota is metered at published overage rates. The math is additive and transparent.

Agora: per-service usage with bundled free tier

Agora gives every new project a 10,000-minute monthly free bucket (expanded August 2025). Beyond that, each service — RTC audio, RTC video, Conversational AI Engine, ARES ASR, Real-Time STT, Cloud Recording, Signaling — is billed on its own per-minute rate. If your app uses four services in one call, you pay four rates.

Self-hosted LiveKit: infrastructure plus operations, no platform fee

The LiveKit server is Apache 2.0. You pay for bare-metal or cloud VMs, bandwidth egress at raw hosting-provider rates (typically $0.01–$0.09 per GB on Hetzner or DigitalOcean, vs $0.10–$0.12 on LiveKit Cloud), and the engineering hours to run it. For AI agents you still pay the upstream LLM/TTS/STT vendors, but nothing to LiveKit itself.

LiveKit Cloud pricing tiers in 2026

LiveKit publishes four tiers plus a la carte overage pricing. Every tier includes WebRTC SFU routing, agent session orchestration, recording, SIP telephony, and the LiveKit Inference gateway for LLM, TTS, and STT calls.

LiveKit overage rates (memorize these)

Agent session minutes: $0.01 per minute above quota. That’s the number you multiply by your projected usage to ballpark a voice-agent bill.

WebRTC participant minutes: $0.0004–$0.0005 per minute (Ship tier and up). Low enough that for most AI-agent apps, you won’t feel it.

SIP telephony: $0.003–$0.004 per minute. PSTN cost is passed through from the carrier separately.

Recording: $0.015 per minute for video, $0.004 per minute for audio-only. Transcoding adds another $0.015–$0.02.

Bandwidth egress: $0.10–$0.12 per GB downstream. Upstream is free. If you’re sending 720p video to 50 receivers per session, egress becomes the biggest line item, not minutes.

Agora pricing in 2026 — RTC plus Conversational AI Engine

Agora sells two different things that get conflated: classical RTC (the original product) and the newer Conversational AI Engine (CAI). If you build an AI voice agent you pay for both, per minute, stacked.

What a typical AI voice-agent minute really costs on Agora

A single minute of an AI voice agent conversation on Agora stacks: CAI Audio Basic ($0.0099) + ARES ASR ($0.0166) = $0.0265. That’s the number on the pricing page. But to actually connect the caller you also consume an RTC audio minute ($0.00099). And if you record the session, that’s another line. Worked example: 1 caller + 1 AI agent in a 10-minute call = ~$0.265 CAI + ~$0.02 RTC + ~$0.005 recording = roughly $0.29 per call before LLM/TTS pass-through.

The v2.5 pricing change (April 2026)

Agora shipped Conversational AI Engine v2.5 on April 1, 2026, which brought updated preset TTS/LLM bundles and a revised pricing schedule. If you signed your Agora MSA before that date, audit your current effective rates — the sales team does not automatically migrate existing customers to new pricing, favorable or not.

The comparison matrix — line by line

This is the matrix clients ask us for when they’re three weeks into evaluation and still can’t decide. It compares the seven axes that actually drive a LiveKit vs Agora decision for an AI-agent product.

Cost model 1 — 100 concurrent voice-AI users, 8 hours per day

Scenario: a mid-market SaaS running an AI customer-support voice line. 100 concurrent sessions, 8 hours per business day, 20 business days per month. That’s 960,000 billable participant-minutes per month. We assume each session uses one human caller and one AI agent, so CAI / agent-session minutes match participant minutes 1:1.

LiveKit Cloud Scale tier

Base plan: $500 per month.
Agent session minutes: 960,000 × $0.01 = $9,600 (quota of 50k absorbs 5%; rounded for clarity).
Participant minutes: 960,000 × $0.0004 = $384.
Inference pass-through (GPT-4o-mini realtime, ~$0.15 per agent-minute): $144,000.
Total: ~$154,500 per month — dominated by inference.

Agora Conversational AI Engine

RTC audio: 960,000 × $0.00099 = $950.
CAI Audio Basic + ARES ASR: 960,000 × $0.0265 = $25,440.
LLM / TTS pass-through (comparable): $144,000.
Total: ~$170,400 per month.

Delta and interpretation

At 100 concurrent users, LiveKit costs roughly $16,000 per month less than Agora — about 10% of the total, 60% of the platform layer. The difference is almost entirely Agora’s $0.0265 vs LiveKit’s $0.0104 platform rate. Inference dominates both invoices; optimizing your system prompt and switching to a smaller model is worth more than switching vendors.

Cost model 2 — 1,000 concurrent users (and where self-hosted wins)

Scale up the same scenario by 10x: 9.6M participant-minutes per month.

At 1,000 concurrent users, the platform-layer gap between LiveKit Cloud and Agora is roughly $163,000 per month, or about $2M per year. Self-hosted LiveKit saves another ~$85,000 per month versus Cloud Scale — but only if your engineering org already handles Kubernetes, multi-region deployment, and 24/7 on-call for media servers. Our rule of thumb: below 500 concurrent sessions, Cloud is cheaper. Above 500, self-hosted earns its keep within two quarters.

Self-hosted LiveKit — when open source pays off

LiveKit Server is Apache 2.0. You can run it on a single $60/month DigitalOcean droplet for prototyping or on a multi-region Kubernetes mesh for production. A 4-core / 8 GB worker comfortably handles 10–25 concurrent AI-agent jobs; a single AWS m5.2xlarge can serve roughly 200 concurrent audio-only participants before you need to horizontally scale.

What self-hosting actually costs

Hosting. On Hetzner AX-series bare metal, bandwidth runs ~$0.001 per GB and a 16-core AX41 machine is €39/month. On AWS with Savings Plans, a reasonable voice-AI production cluster (3–4 m5.2xlarge + Redis ElastiCache + NLB) runs $1,500–$3,500/month for 1,000 concurrent sessions.

Engineering. Plan on 0.5–1.0 FTE of SRE time ongoing. That’s the number most build-vs-buy spreadsheets forget. At a conservative $150k loaded cost per engineer, that’s $75k–$150k per year — real money, but still a rounding error next to $2M of platform fees at scale.

Time to first production traffic. Using Agent Engineering we ship a multi-region self-hosted LiveKit cluster with monitoring, autoscaling, and SIP ingress in 4–6 weeks for most clients. Without AI-assisted tooling, plan for 10–12. See our LiveKit AI agent development service for how we typically scope these engagements.

When we recommend self-hosting

Consistent concurrency above 500 sessions, data residency requirements Cloud can’t hit, or a unit-economics model where gross margin is below 60% and platform cost is the next lever. Otherwise, stay on Cloud. The time you save going to market almost always beats the margin you’d recover.

Feature parity — what each platform does better

Pricing is only the headline. The feature set drives a lot of decisions, because the “5% cheaper” platform that forces you to build interruption handling yourself is never actually cheaper.

LiveKit’s strengths

1. Pluggable inference. The livekit-agents Python and Node frameworks let you swap OpenAI, Anthropic, Groq, Deepgram, ElevenLabs, or your own fine-tuned model behind the same interface. You’re never locked to a vendor’s bundled STT.

2. First-class telephony. LiveKit ships both inbound and outbound SIP, plus its own phone-numbers product. The April 2026 Telnyx partnership brought sub-200 ms PSTN latency and carrier-grade SIP at roughly 50% lower cost versus rolling your own.

3. Compliance on demand. SOC 2 Type II is table-stakes, but the HIPAA BAA on Scale tier is what matters if you’re in telehealth, insurance, or financial services. Agora still requires an enterprise contract for a BAA.

Agora’s strengths

1. SD-RTN is one of the largest private RTC networks on Earth. 200+ countries, measured 33 ms intra-regional North America vs 94 ms over public internet. If your traffic is heavy in Southeast Asia, India, or Latin America, Agora’s edge wins.

2. Agent Studio for non-developers. Agora’s low-code Agent Studio lets product managers configure ASR/LLM/TTS presets and deploy agents without touching SDK code. Useful for contact-center pilots where design iteration matters more than model control.

3. Massive SDK footprint. Agora has native SDKs for Unity, Unreal, Flutter, React Native, Electron, Cocos, and every mobile platform older than 2018. For game or XR studios that matters.

Hidden costs and gotchas we’ve hit in production

These are the surprises in a real LiveKit vs Agora pricing comparison — the ones that show up on invoice three, after management has already committed to the platform. Budget for all of them.

Agora-specific

1. RTC + CAI double-billing. A 1-minute AI-agent call doesn’t cost $0.0265. It costs $0.0265 + the underlying RTC audio minute + any recording multiplier.

2. Recording billed N × M. If five participants are recorded for one minute, that’s 5 service-minutes on your invoice. Easy to forget on a live group class or a multi-party support bridge.

3. Real-Time STT standby fee. If STT is enabled but all hosts are muted, you still pay $0.99 per 1,000 minutes. Teams that leave STT “on by default” in quiet classrooms burn thousands of dollars a month on silence.

LiveKit-specific

1. Bandwidth egress to S3. Exporting recordings to S3/GCS costs the same $0.10–$0.12 per GB as streaming them to clients. A 720p 1-hour recording is ~900 MB; at 10k recordings per month, that’s $900–$1,100 just to write to your own bucket.

2. Inference credits burn fast with large models. LiveKit Inference bills pass-through rates. Claude 3.5 Sonnet at 1,500 tokens per turn, 20 turns per conversation, costs about $0.09 per conversation — $5 of free credits lasts ~55 conversations. Use cheaper models for drafts; reserve Sonnet/Opus for final agents.

3. SIP minutes aren’t in the agent minute quota. PSTN is billed separately. If you pivot from browser to phone traffic mid-quarter, re-forecast.

A decision framework — pick the right platform in five questions

Run these five questions in order. Answer “yes/no/unsure”. If four out of five point to one platform, go. If it’s 3:2, go back to the cost model.

Q1. Is your app an AI agent, a pure video/voice calling feature, or both? AI-agent-first → LiveKit. Pure RTC → Agora RTC or self-hosted. Both → LiveKit, because RTC on LiveKit is cheap enough and you want one vendor for the agent layer.

Q2. Where do your users live? North America and EU → LiveKit’s mesh is competitive. Southeast Asia, India, or Latin America at scale → Agora SD-RTN still has the edge for pure RTC.

Q3. Do you need HIPAA, SOC 2 Type II, or data residency? Yes → LiveKit Scale/Enterprise ships HIPAA BAA and SOC 2 off the shelf. Agora requires a dedicated enterprise contract.

Q4. Who owns the inference model choice — you or the platform? You want to A/B test OpenAI vs Anthropic vs a fine-tuned Llama → LiveKit’s plugin system. You want a bundled, opinionated stack that just works → Agora Agent Studio.

Q5. Is your engineering team ready to run media servers? Yes, with SREs on call → consider self-hosted LiveKit above 500 concurrent. No → stay on Cloud. “Maybe” almost always resolves to no once you price 24/7 rotations.

Mini case — how we scoped a 600-seat AI receptionist for a fitness chain

Situation. A fitness chain operating 80 clubs wanted a 24/7 AI receptionist to handle membership inquiries, class bookings, and member-retention calls. Expected load: ~600 concurrent calls at peak (Saturday mornings, pre-class), ~50,000 conversations/month averaging 4 minutes. They had been quoted roughly $45,000/month on Agora CAI through a regional systems integrator.

Plan (12 weeks, Agent-Engineering-assisted). Weeks 1–2: requirements, knowledge-graph from their booking system, LLM prompt harness. Weeks 3–6: LiveKit Cloud Scale deployment, OpenAI Realtime + Deepgram Nova-3 STT, SIP ingress through Telnyx. Weeks 7–9: interruption tuning, escalation-to-human flows, cadence to their Twilio Flex instance. Weeks 10–12: soak test at 1.5x projected peak, compliance review, cutover.

Outcome. Platform cost landed at $19,400/month (LiveKit Scale + inference) vs the $45,000 Agora quote — 57% savings. First-call resolution measured at 71% after six weeks in production. Average turn-taking latency: 420 ms end-to-end. Our scoping fee, including migration-risk contingency, paid back inside the first billing cycle. Want a similar assessment? See also our AppyBee fitness booking platform case for context on our fitness-vertical experience.

Five pitfalls to avoid during platform selection

1. Forecasting with list prices only. Neither vendor’s published rates include your LLM pass-through, your bandwidth, or your recording multiplier. Any forecast that doesn’t show inference and egress as separate line items is wrong.

2. Confusing RTC minutes with agent minutes. They are different billable units. A 60-minute AI-agent call is 60 agent minutes AND 60 RTC participant minutes on Agora, and it’s 60 agent-session minutes AND 60 WebRTC minutes on LiveKit. Single-unit math underestimates cost by 15–30%.

3. Assuming enterprise discounts will save you. They sometimes do. But “we’ll get a 40% discount at scale” is a hypothesis, not a plan. Get it in writing before you commit your stack, especially on Agora where the quoted discount tier can swing 20% based on your contracting officer.

4. Underestimating migration cost. Moving an existing WebRTC app from Agora to LiveKit is not a weekend project. Expect 4–8 weeks for a typical small-to-mid SaaS, double that if you rely on Agora’s proprietary signaling (RTM). Factor the engineering cost into your break-even calculation.

5. Over-engineering for unlikely geographic edge cases. If 3% of your users are in Jakarta, you don’t need SD-RTN — you need a regional TURN server. Don’t buy the wrong platform for the edge case.

KPIs to measure after you pick

Quality KPIs. End-to-end turn-taking latency p50 under 600 ms and p95 under 1,200 ms. Interruption recovery under 200 ms. Word error rate under 8% on domain-specific vocabulary. Hold these thresholds or your users will call the agent “slow and dumb” regardless of vendor.

Business KPIs. Cost per completed conversation, containment rate (conversations resolved without human handoff), and first-call resolution. A good voice-AI product lands containment at 55–75% and cost per conversation under $0.30 inclusive of inference.

Reliability KPIs. Session setup success rate above 99.5%. Mean time to recovery on a regional outage under 90 seconds. Track them per-region; the global average hides the bad geography.

When to not pick either platform

Both LiveKit and Agora are WebRTC-based SFUs with AI orchestration bolted on. That’s the right stack for 90% of AI-agent and video-calling use cases. But not all.

If you’re building an asynchronous AI product — text chatbot, email responder, batch call summarization — you don’t need WebRTC at all. A serverless function plus an LLM API plus a queue is cheaper and simpler. Don’t pay for real-time infrastructure you won’t use.

If your product is broadcast-style live video (10,000+ viewers, one-to-many), HLS/DASH over a CDN is typically 10–20x cheaper than SFU-based RTC. See our live streaming platform cost breakdown for the economics.

If you need a fully open ecosystem with no vendor managed service at all, look at mediasoup, Janus, or Kurento. Expect higher total engineering cost but zero platform lock-in. Our team has shipped production deployments on all three, and we discuss when each wins in P2P vs MCU vs SFU.

FAQ

What to read next

Ready to escape per-minute surprise invoices?

Here’s the one-paragraph LiveKit vs Agora pricing verdict. For AI-agent workloads, LiveKit almost always lands cheaper than Agora on the platform layer — about $0.01 vs $0.0265 per minute — and the delta grows with scale. For pure RTC at global scale in APAC or LATAM, Agora’s SD-RTN still has a real edge. Self-hosted LiveKit saves another 30–50% once you clear ~500 concurrent sessions, but only if you can staff it. Most hidden costs aren’t in the per-minute rate; they’re in bandwidth egress, recording multipliers, and inference pass-through. Model those before you sign.

If you’re three weeks into this evaluation and still not sure, bring us your invoices and your peak concurrency. We’ll run the math, show you the break-even, and tell you whether to switch, stay, or self-host. No opinion without your numbers.