ThunderPhone vs. LiveKit Cloud

ThunderPhone and LiveKit Cloud both put AI voice agents on real phone calls, but they sit at different altitudes. LiveKit Cloud is managed infrastructure: it hosts, scales, and connects voice agents, but the agent itself is a program you write, deploy, and operate on it. ThunderPhone is a finished hosted phone-agent product: the agent already exists, and you configure it. Neither is a lesser version of the other, and plenty of strong engineering teams should pick LiveKit Cloud on purpose.

This page compares ThunderPhone against LiveKit's managed offering: LiveKit Cloud running agents built with the open-source LiveKit Agents framework. If you are weighing a fully self-hosted stack instead, that is a different decision, covered in ThunderPhone vs. building your own voice-agent stack. Every LiveKit fact below was checked against LiveKit's public site and documentation on August 19, 2026; ThunderPhone facts reflect its published pricing and documentation as of the same day.

Decision summary: choose LiveKit Cloud if you have engineers who want to own the agent code — pick every STT, TTS, and LLM and control each layer — while delegating realtime infrastructure and scaling to a managed platform. Choose ThunderPhone if you want a phone agent in production without writing or operating agent code, at a published 2¢/5¢/9¢ per-minute rate with languages, telephony, transfers, knowledge, webhooks, and testing already built in.

At a glance: what's managed, what remains yours

Cost scenario: 10,000 connected minutes per month

Use one explicit workload: 2,000 calls × 5 connected minutes = 10,000 connected minutes per month, one U.S. phone number, calls over a third-party SIP trunk, one included primary language, no hold, voicemail, or transfer minutes. This is illustrative usage arithmetic, not a quote.

LiveKit Cloud: the infrastructure subtotal is quotable; the all-in total is not

LiveKit Cloud's infrastructure components have published rates, so that layer is computable. On the Ship plan ($50/month):

• Plan fee: $50
• Agent session minutes: 5,000 included, then 5,000 × $0.0100 = $50
• Third-party SIP minutes: 5,000 included, then 5,000 × $0.004 = $20
• Phone number: first US local number free, so $0
• Infrastructure subtotal: ~$120/month. LiveKit pricing, accessed 2026-08-19.

On the Scale plan ($500/month), the 50,000-minute agent-session and SIP allowances absorb the whole workload, so the infrastructure subtotal is the $500 plan fee.

The layer that cannot be computed from public pages is the model layer. Ship includes $5 of LiveKit Inference credit — which LiveKit equates to roughly 100 minutes "based on model prices" — and Scale includes $50 (~1,000 minutes) with discounted rates thereafter; beyond the credits, inference bills per model, or you pay providers directly through plugins. For a 10,000-minute month, most inference is billed by whichever STT, TTS, and LLM combination you configured. A configured all-in total for this workload is therefore not established from the reviewed public sources — it is the infrastructure subtotal above plus a model bill that depends on your choices. For teams who want to pick and negotiate every model, that per-component visibility is a feature, not a gap. LiveKit pricing and models docs, accessed 2026-08-19.

ThunderPhone: the engine rate is the call price

• Spark: 10,000 × $0.02 = $200/month
• Bolt: 10,000 × $0.05 = $500/month
• Storm: 10,000 × $0.09 = $900/month
• Subscription or seat fee: $0. A premium voice or language path could add up to 10,000 × $0.03 = $300/month; telephony and other optional configuration charges depend on setup.

Hold time bills a flat 2¢/minute, and voicemail and phone-menu minutes bill at the engine rate; the scenario assigns zero minutes to those states. Only the base engine usage is computable from minutes alone — but it is computable, from public numbers, before any build starts.

The two results are not the same kind of number. ThunderPhone's $200–$900 is a configured product's usage bill; LiveKit Cloud's ~$120–$500 is an infrastructure subtotal beneath an agent you still have to build, with model costs and engineering time on top.

What LiveKit Cloud manages — and what you still own

LiveKit Cloud genuinely removes the hardest infrastructure work from a voice-agent build. Deployment is "a single LiveKit CLI command"; the platform runs agents on its global network with "automatic scaling and load balancing," builds container images from your Dockerfile, stores secrets, exposes logs, and supports rollouts and rollbacks. The Agents framework underneath handles "streaming audio through an STT-LLM-TTS pipeline, reliable turn detection, handling interruptions, and LLM orchestration." LiveKit deployment docs and Agents docs, accessed 2026-08-19.

What remains yours is the agent itself:

• The code. Your agent is a Python or Node.js program; prompts, tool calls, state handling, and every business rule are software you write and maintain. LiveKit Agents docs, accessed 2026-08-19.
• Model selection and its costs. You choose the STT, TTS, LLM, or realtime speech-to-speech model — via LiveKit Inference or your own provider accounts — and own the resulting tuning and bills. LiveKit models docs, accessed 2026-08-19.
• Telephony configuration. Trunks, dispatch rules, and number provisioning are yours to set up. LiveKit SIP docs, accessed 2026-08-19.
• Testing. Every test case, judged intent, simulated conversation, and pass criterion is authored by your team. LiveKit testing docs, accessed 2026-08-19.
• Operations above the platform line. Dockerfile, secrets, rollouts, rollbacks, and monitoring interpretation stay with you.

ThunderPhone draws the line higher: the platform owns everything on that list, and your surface is configuration — prompts, 47 languages and 123 voices with automatic mid-call switching, cold and attended warm transfer, a knowledge base with per-agent scoping and mid-call search, HMAC-SHA256-signed webhooks, remote MCP servers over streamable HTTP, an embeddable web widget, and a built-in testing suite. The trade is symmetric: you cannot swap the STT model or read the pipeline code, and in exchange none of it is yours to build, test, or operate.

Choose LiveKit Cloud if

• You have engineers who want to own the agent code and pick every model — specific STT, TTS, LLM, or realtime speech-to-speech choices, swappable at will. Teams in this position should strongly consider it. LiveKit models docs, accessed 2026-08-19.
• Your agent needs custom pipeline behavior — bespoke turn handling, non-phone modalities, or logic that only makes sense as your own code — on managed infrastructure rather than self-run servers. LiveKit Agents docs, accessed 2026-08-19.
• You want per-component cost visibility: published infrastructure rates plus model bills you select and negotiate yourself. LiveKit pricing, accessed 2026-08-19.
• Your procurement requires a stated SOC 2 Type II posture at the infrastructure layer, with HIPAA BAAs available on Scale and Enterprise plans. LiveKit security page, accessed 2026-08-19.
• Voice agents are core product, and investing engineering headcount in the agent layer is deliberate strategy.

Choose ThunderPhone if

• You want a phone agent in production without writing, deploying, or operating agent code, at a price you can read before starting: 2¢/5¢/9¢ all-in engine rates, pay as you go, no subscription or seat fee.
• Multilingual coverage matters and you would rather configure from a published catalog of 47 languages and 123 voices with mid-call switching than select and validate model combinations in code.
• You need call features that are expensive to build well — attended warm transfer with private briefing and fallback, voicemail policy handling, phone-menu navigation, mid-call knowledge search.
• You want release-grade testing as a product feature — graded scenarios, regression gates, CI execution, live A/B — rather than a test suite your team authors and maintains.
• GDPR and HIPAA cover the public compliance posture you need, with your team handling its own configuration obligations.

FAQ

Is LiveKit Cloud a finished voice-agent product?

No — it is managed infrastructure for agents you build. LiveKit Cloud hosts, scales, and connects voice agents written with the LiveKit Agents framework, but the agent is a Python or Node.js program your team writes, deploys, and operates; business logic is the developer's responsibility. ThunderPhone is the opposite shape: a finished hosted agent you configure. LiveKit Agents docs, accessed 2026-08-19.

What does 10,000 minutes per month cost on each?

ThunderPhone's base-engine arithmetic is $200–$900; LiveKit Cloud's infrastructure subtotal is roughly $120 (Ship) to $500 (Scale), with the model layer on top. A configured LiveKit all-in total depends on which STT, TTS, and LLM you choose and is not established from the reviewed public sources. Both figures are illustrative arithmetic, not quotes. LiveKit pricing, accessed 2026-08-19.

Can both handle real phone calls over SIP?

Yes, by different routes. LiveKit Cloud supports bring-your-own third-party SIP trunks with dispatch rules, and sells US local and toll-free numbers. ThunderPhone provides inbound demo numbers for testing plus direct VoIP connections or manual SIP configuration for any SIP trunk, with verified imported numbers supporting inbound and outbound calling. LiveKit SIP docs, accessed 2026-08-19.

How do the compliance postures differ?

LiveKit publicly states SOC 2 Type II, GDPR alignment with a DPA, HIPAA BAAs on Scale and Enterprise plans, CCPA/CPRA, and the EU-US Data Privacy Framework; ThunderPhone's public posture is GDPR and HIPAA compliant. On LiveKit Cloud those labels cover the infrastructure — the agent code and any directly-integrated providers carry postures you must assess separately. On either product, customers remain responsible for configuring workflows, data handling, recordings, and notices for their own obligations. LiveKit security page, accessed 2026-08-19.

Who writes and runs the tests?

On LiveKit Cloud, your team does, with good tooling; on ThunderPhone, the suite is a product feature. LiveKit Agents ships pytest/Vitest behavioral testing with LLM-judged assertions and cloud-run simulated conversations, all authored by your developers. ThunderPhone provides AI-caller simulations, SIP-loopback calls, graded scenarios, regression gates, CI execution, and live A/B experiments as configuration; simulations are billable real calls. LiveKit testing docs, accessed 2026-08-19.

Sources and freshness

All LiveKit claims on this page were checked against livekit.com and docs.livekit.io on August 19, 2026, with per-claim source links shown throughout. ThunderPhone pricing and product details reflect its published pricing and documentation as of August 19, 2026. This comparison is documentation-based — no LiveKit Cloud deployment was built or billed — and vendors change pricing and documentation often, so confirm against the live pages before relying on them.