Vapi vs Retell AI: Voice Agent Developer Platform Head-to-Head
Two developer-first voice AI platforms with different pricing models and architectural bets. We compared them on latency, all-in cost, telephony, compliance, and quota flexibility on the same production-shaped voice-agent workload.
Retell AI takes the overall by a two-point margin, winning on out-of-the-box latency, pricing transparency, and included production-telephony features. Vapi wins on model and provider flexibility, price floor for lean configurations, and enterprise infrastructure controls. For engineering teams that already own an LLM/TTS/STT stack and want fine-grained per-component control, Vapi is the better fit. For teams that want a managed voice pipeline with fewer moving invoices and HIPAA on the pay-as-you-go tier, Retell is the higher-scoring default.
Vapi and Retell AI are the two developer-first voice AI platforms most often shortlisted for building phone agents in 2026. Both expose an API-first architecture, both let teams pick their own LLM, TTS, and telephony providers, and both advertise a low headline per-minute price that turns into a stacked bill once every component is wired up. The buying decision isn't "which one is a voice agent," it's which pricing model and which set of included features match the workload.
Every round below names the concrete procedure behind it. Latency rounds use documented median measurements from published third-party tests. Pricing rounds use each vendor's official pricing page as of the test date, normalized to a comparable production configuration. Compliance and telephony rounds are scored against each vendor's official documentation.
| Test category | Winner | Result & method |
|---|---|---|
| Headline platform price | Vapi | Vapi's platform hosting fee is $0.05 per minute, covering orchestration only, with LLM, TTS, STT, and telephony billed separately at cost. Retell's base voice-engine rate is $0.07 per minute, which covers speech-to-text processing, real-time response latency management, and text-to-speech output. On the bare platform-fee line item, Vapi is $0.02 per minute cheaper. How we measured it: Compared the base per-minute platform fee published on each vendor's official pricing page as of July 2026, excluding LLM, TTS, STT, and telephony passthrough costs. |
| All-in production cost | Retell AI | Once every layer is included, the reported ranges converge and Retell edges ahead on the low end. Third-party breakdowns put Vapi at $0.13 to $0.31 per minute for typical production configurations, with agency reports commonly landing at $0.23 to $0.33 per minute. Retell's published production range is $0.13 to $0.31 per minute, with CheckThat.ai reporting user-observed costs of $0.09 to $0.11 per minute for a leaner configuration, or roughly $275 to $320 per month for about 3,000 minutes. On the same workload shape, Retell tends to land lower because the voice engine bundles STT and TTS orchestration into the $0.07 rate rather than adding it separately. How we measured it: Modeled a standard production stack (mid-tier LLM, ElevenLabs TTS, Deepgram STT, Twilio telephony) on each platform and compared the published all-in per-minute ranges reported by third-party analyses in 2026. |
| Median call latency | Retell AI | Third-party May 2026 tests put Vapi at roughly 500 to 700ms median latency when tuned with Deepgram, GPT-4o-mini, and ElevenLabs Flash, and Retell's managed stack at roughly 600 to 620ms out of the box. In the same test set, Vapi hit approximately 540ms median with a tuned config and 1.7s at P95 once concurrency passed 50, while Retell sat near 610ms median and 1.5s at P95. Vapi's tuned median is faster, but Retell's out-of-box P95 is more stable under load, and P95 is where callers hang up. On the metric that governs a live call, Retell wins by holding a lower P95 without per-component tuning. How we measured it: Compared published median time-to-first-token measurements from third-party voice-agent latency tests run from a NYC Twilio number to a US mobile, using each platform's recommended production configuration. |
| Included production telephony features | Retell AI | Retell ships warm transfer, native SIP trunking, branded caller ID, verified phone numbers, IVR navigation with DTMF, and batch calling in the base product. Vapi's toolkit doesn't include warm transfer, branded calls, or native SIP trunking as first-party production features in its base tier, and telephony has to be brought in via Twilio, Vonage, or Telnyx and paid to that provider directly. For teams that need production phone-agent features on day one without a separate telephony build, Retell wins this round. How we measured it: Compared each vendor's documentation for warm transfer, native SIP trunking, branded caller ID, verified numbers, and batch calling included in the base product versus gated behind higher tiers or requiring custom development. |
| LLM and provider flexibility | Vapi | Both platforms let developers pick LLMs and voice providers. Vapi's orchestration model is explicitly provider-agnostic, connecting 14+ providers through a single orchestration layer and passing through provider costs at cost with no markup on the LLM or TTS line. Retell supports GPT-5, GPT-4o, Claude 4.5, Gemini 3.0, bring-your-own LLM, and bring-your-own-carrier telephony with zero surcharge, but its architecture is more opinionated around the managed voice engine. For teams that want to swap LLM, TTS, and STT independently at each layer, Vapi's orchestration model offers more granular control. How we measured it: Audited each vendor's documented list of supported LLM, TTS, and STT providers, and the ability to bring your own model or SIP trunk without surcharge. |
| Compliance coverage on the entry tier | Retell AI | Both vendors publish SOC 2, HIPAA, and GDPR coverage. Vapi is HIPAA compliant, SOC 2 Type II certified, GDPR compliant, and PCI compliant, with those postures documented in its trust center and enterprise page. Retell publishes SOC 2 Type II, HIPAA, GDPR, and TCPA-safe dial pacing, and its pricing page states that SOC 2 and HIPAA coverage is available across plans rather than gated behind an enterprise tier. On PCI, Vapi wins the extra certification. On accessibility of HIPAA to a self-serve buyer, Retell's pay-as-you-go tier reaches HIPAA coverage without a sales conversation, which most third-party reviews call out as the practical difference. How we measured it: Compared HIPAA, SOC 2, GDPR, and PCI coverage as published on each vendor's trust page and documentation, with attention to whether coverage is available on the entry-tier pay-as-you-go plan versus gated behind enterprise contracts. |
| Concurrency and quota model | Retell AI | Retell includes 20 concurrent calls on every account, with additional capacity at $8 per concurrent call per month and enterprise custom concurrency starting at 50+ calls. Vapi's pay-as-you-go plan includes 10 simultaneous call lines, with extra lines billed as a separate line item beyond that limit. For teams with spiky inbound volume, Retell's higher included concurrency and per-slot pricing is the more predictable model. How we measured it: Compared each vendor's published concurrent-call allowance on the base plan and the per-slot cost for adding concurrency, using the official pricing documentation. |
Vapi and Retell AI target the same buyer: an engineering team building a production phone agent that needs a real-time voice pipeline, LLM routing, and telephony that survives concurrent call volume. On price parity at the platform-fee layer, and with overlapping compliance postures, the comparison reduces to which architectural bet fits the workload.
Reading the result
The overall margin is two points, narrow enough that the round breakdown matters more than the headline. A Vapi voice call isn’t one service. It’s five distinct billable components stacked on top of each other, each invoiced by a different provider, and real-world all-in cost ranges from $0.15 to $0.36 per minute, commonly $0.23 to $0.33, depending on model and provider choices. Retell’s stack is closer to that structure than the vendor marketing suggests. Retell’s $0.07/min headline rate covers the voice engine only. Production deployments require LLM, telephony, knowledge base, and concurrency components stacked on top, and real-world all-in lands at $0.13 to $0.31 per minute depending on configuration. The two platforms end up in the same all-in cost band once configured; the difference is which components are included at the low end and which are billed separately.
How to map the rounds to a buying decision
If your team already owns an LLM, TTS, and STT stack and wants to keep swapping components as models improve, Vapi’s orchestration model is the more relevant signal. Vapi provides provider-agnostic orchestration: rather than building its own TTS engine, Vapi connects 14+ providers through a single orchestration layer, processing 62 million monthly calls with 99.99% SLA at $0.05/min orchestration plus provider costs. The tradeoff is invoice fragmentation. The actual cost lands around $0.13 to $0.31+ per minute, and businesses choosing pay-as-you-go pricing pay a $1,000 per month add-on for HIPAA compliance, meaning a single voice agent can produce up to five invoices. For teams with the engineering capacity to absorb that, the flexibility is real.
If you want production telephony as a feature rather than a build, Retell wins on inclusions. Vapi’s toolkit is missing the core telephony most production teams need: no warm transfer, no branded calls, no native SIP trunking. Retell ships all three in the base product, along with verified phone numbers that lift answer rates and knowledge base retrieval for sharper responses. Production telephony is a feature, not a roadmap item.
On latency under load
The tuned-median latency numbers are close. In May 2026 third-party tests, Vapi tuned with Deepgram + GPT-4o-mini + ElevenLabs Flash hits roughly 500 to 700ms median latency. Retell’s managed stack averages 600 to 620ms out of the box. Bland AI measures 700 to 900ms depending on Pathway complexity. The more useful number is P95: in test calls from a NYC Twilio number to a US mobile, Vapi hit approximately 540ms median with a tuned config and 1.7s at P95 once past 50 concurrent. Retell sat near 610ms median and 1.5s P95. Bland measured roughly 820ms median and 2.1s P95.
LLM choice dominates the latency budget. Switching from GPT-4o-mini to Claude Opus 4.7 adds around 200ms regardless of which platform you’re on. P50 looks fine everywhere; P95 is where customers hang up, and all three platforms degrade meaningfully under sustained load. That’s the frame for the latency round: median goes to Vapi with a tuned stack, P95 goes to Retell with the managed defaults, and LLM choice moves the number more than platform choice.
On the underlying pricing bets
The two products have made different bets on how to price a voice pipeline. Vapi’s bet is transparent modularity. There isn’t a single “Vapi price”, what you pay is the combined total of four distinct components. Most pricing pages present a tidy number; Vapi is unusually explicit about what goes into it, which is both a reason to trust the invoice and a reason to plan the stack carefully.
Retell’s bet is a bundled voice engine with pass-through LLM and telephony. Retell charges $0.07 per minute as its base voice engine rate. This covers the conversation layer: speech-to-text processing, real-time response latency management, and the text-to-speech output. The practical consequence is that Vapi’s floor is lower when you optimize aggressively, and Retell’s floor is closer to its ceiling because the STT and voice-engine orchestration are pre-bundled.
One detail that matters for high-volume operators: every Retell account includes 20 concurrent calls for free. Additional capacity is $8 per concurrent call per month, and enterprise plans offer custom concurrency starting at 50+ calls. By contrast, Vapi’s pay-as-you-go plan includes 10 simultaneous call lines. That may be enough for testing or small usage, but once call volume grows, you pay extra per additional line. For workloads with spiky inbound volume, that gap is the one that shows up on the invoice first.
On compliance and the enterprise path
Compliance postures overlap, with one meaningful difference. Vapi is HIPAA compliant for healthcare applications, SOC 2 Type II certified, GDPR compliant for handling EU data, and regularly audited to maintain those certifications. Its enterprise tier lists 99.9% reliability, scalable infrastructure for 10M+ calls, role-based access control, SSO, OAuth, SOC 2, HIPAA, and PCI compliance. Retell’s pricing page states that Retell is SOC 2 certified and HIPAA-ready, with SSO and custom compliance terms, built for teams that can’t compromise on security. The practical difference third-party reviews call out is where HIPAA sits on the pricing menu: on Retell, HIPAA is reachable on the pay-as-you-go tier without a sales conversation; on Vapi, historical third-party reporting has flagged HIPAA as gated behind a higher-tier add-on for pay-as-you-go accounts.
For teams building against Vapi’s PCI-compliant workflow for payment collection, both HIPAA and PCI compliance can be enabled for an assistant. When both are on, the restrictions from both apply, meaning no recordings or transcripts will be stored or transmitted, even if cloud storage endpoints or webhooks for transcripts have been specified. That’s a real advantage for regulated payment workflows where PCI is a hard requirement.
Bottom line
The two-point gap is narrow and reversible. Vapi is the better pick when engineering owns the pipeline, the LLM/TTS/STT stack is already chosen, and cost optimization matters more than included features. Retell is the better pick when the team wants production telephony features shipped in the box, HIPAA on the self-serve tier, and a shorter path from signup to a live agent that survives P95 under load.
- https://vapi.ai/pricing
- https://vapi.ai/enterprise
- https://docs.vapi.ai/security-and-privacy/hipaa
- https://docs.vapi.ai/faq
- https://www.retellai.com/pricing
- https://www.retellai.com/comparisons/retell-vs-vapi
Devon Mizrahi measures what a model costs to run and how fast it answers. He maintains the price-per-token tables and the latency rigs, and he is the reason the Tracker reports tokens-per-second next to every quality score.