If you're building an AI voice agent, your first decision is whether to use a hosted platform or build the telephony/STT/TTS/LLM stack yourself. The hosted platforms — Vapi, Retell, and Bland — have all matured significantly in the past year, but they're optimized for different use cases. Getting this choice wrong wastes 2–4 months.
This post gives you a straight comparison across the dimensions that actually matter: latency, pricing, HIPAA compliance, telephony flexibility, and how much control you get over the conversation logic.
What the platforms do (and don't do)
All three platforms — Vapi, Retell, Bland — abstract the following away from you:
- WebSocket audio pipeline (capturing microphone, streaming to STT, streaming TTS back)
- Telephony integration (they provision numbers and handle the SIP trunk)
- Turn management (knowing when the caller has finished speaking)
- Barge-in handling (interrupting the agent when the caller starts talking)
What you still have to build yourself on any platform:
- Conversation design and system prompts
- Business logic (CRM lookups, calendar bookings, database writes)
- Call routing logic
- Post-call processing (summaries, CRM updates, follow-up tasks)
The platforms handle the plumbing. You handle the intelligence.
Vapi
Vapi launched in late 2023 and has become the default choice for developers building voice agents for the first time. The developer experience is genuinely good — you can have a working agent on a real phone number in under an hour.
Pricing: $0.05/minute for the platform layer (audio pipeline, telephony), plus your LLM costs. A 3-minute call costs $0.15 in platform fees + whatever you pay OpenAI or Anthropic. At 1,000 calls/month: $150 in platform fees + $10–$80 in LLM costs depending on model.
Latency: End-to-end response latency (caller stops speaking → agent starts responding) is typically 1,000–1,400ms with GPT-5. Using GPT-5 mini or Groq can bring this down to 800–1,000ms. Vapi introduced turbo mode which helps, but they're still behind Retell on raw latency.
LLM support: Works with any LLM via OpenAI-compatible API. You can use OpenAI, Anthropic, Google, Groq, or a self-hosted model. This flexibility is a genuine advantage over Bland.
Telephony: Vapi manages numbers for you (Twilio under the hood) or you can bring your own Twilio SID. SIP trunk support is available for enterprise customers.
HIPAA: Not HIPAA-compliant on the standard plan. Vapi is working on it, but as of mid-2026 it's not production-ready for PHI. If you're in healthcare, this is a dealbreaker.
Custom logic: Vapi supports tool use — you define functions the agent can call (lookup order status, book a slot, check inventory) and Vapi handles the function calling. Complex logic can be handled via webhooks to your server. Works well for moderate complexity; breaks down for highly stateful or branching conversation flows.
Best for: Fast time to market for non-HIPAA use cases. Sales development, lead qualification, appointment reminders, customer support.
Retell AI
Retell is the performance-focused platform. Their core differentiator is latency — they've invested heavily in minimizing the gap between "caller stops speaking" and "agent starts responding."
Pricing: $0.04–$0.07/minute depending on plan and LLM, inclusive of STT/TTS. Custom pricing available for volume. At 1,000 calls averaging 3 minutes: $120–$210/month inclusive.
Latency: 700–1,000ms end-to-end with their optimized stack. This is the lowest latency of the three hosted platforms. On phone calls, 700ms feels like natural conversation; 1,400ms starts to feel like a robocall. If you're building a sales or customer service agent where the interaction feel matters, Retell's latency advantage is real.
LLM support: OpenAI-compatible API models. Retell has tested and optimized prompts for GPT-5, but other models work with configuration.
Telephony: Similar to Vapi — managed numbers or bring your own. SIP trunk support available.
HIPAA: Retell offers a HIPAA-compliant option for enterprise customers with a signed BAA. This is a meaningful differentiator over Vapi for healthcare use cases. Expect pricing negotiations for the enterprise tier.
Custom logic: Retell supports LLM function calling and webhooks. Their "state machine" conversation design is useful for structured conversations with defined paths (intake forms, screening questionnaires). Less flexible for highly dynamic, unscripted conversations.
Best for: Customer-facing applications where conversation feel matters. Healthcare (with their enterprise HIPAA tier). Structured intake flows.
Bland AI
Bland is the simplest and cheapest of the three platforms, designed for high-volume outbound campaigns and simple call flows.
Pricing: $0.09/minute flat, with LLM included. This is higher per-minute than Vapi or Retell, but includes the LLM cost and is simpler to budget. At 1,000 calls of 3 minutes: $270/month all-in.
Wait — that's more expensive than Vapi. The simplicity isn't in the price, it's in the setup. Bland has the fastest "time to first call" of the three platforms. If you need 10,000 outbound calls running in 48 hours, Bland is designed for that.
Latency: 1,000–1,400ms, similar to Vapi. Not Retell.
LLM support: Bland uses their own model selection internally. You have less control over which exact model runs your conversation. This means you can't swap to Groq for cost savings or Claude for specific capabilities.
Telephony: Strong outbound dialing infrastructure — Bland handles the dialer, retry logic, and voicemail detection for outbound campaigns. Better than Vapi/Retell for bulk outbound.
HIPAA: Not HIPAA-compliant. Not appropriate for PHI.
Custom logic: More limited than Vapi or Retell. Better for scripted flows than dynamic conversations.
Best for: High-volume outbound campaigns (appointment reminders, surveys, lead nurturing at scale). Simple, scripted inbound flows. When you need calls running fast and don't need high flexibility.
Custom build
A custom build means you control every component: choose your telephony provider, your STT model, your TTS voice, your LLM, and how they're wired together.
Operating cost: $0.04–$0.06/call at 1,000 calls/month, $0.03–$0.05 at 10,000+ calls/month. Significantly cheaper than any hosted platform at scale.
Latency: You can match or beat Retell with the right component choices. Deepgram Nova-3 for STT (sub-300ms streaming latency), Cartesia Sonic or Deepgram Aura for TTS (30–50ms TTFB), Groq for LLM inference (100–300ms). End-to-end is achievable at 600–900ms with tuned infrastructure.
HIPAA: Fully configurable. You choose every provider and sign BAAs with each. Self-hosted models and self-hosted vector databases mean PHI never leaves infrastructure you control.
Custom logic: Total flexibility. Complex business logic, stateful conversation management, multi-turn tool use, dynamic persona switching, branching flows — all implementable. No platform limitations.
Build cost: 150–300 hours for a production-ready system. At $100–$150/hour: $15,000–$45,000. See the detailed cost breakdown in our AI voice agent cost guide.
What you're taking on: WebSocket audio pipeline stability, STT streaming error handling, audio buffer management, silence detection tuning, barge-in logic, infrastructure monitoring, and all the DevOps that comes with running distributed real-time systems.
Best for: Healthcare (HIPAA required), high call volume (>10K/month where cost savings justify build cost), complex business logic, building a platform (your product is the voice agent), or situations where vendor dependency is a risk.
Feature comparison matrix
| Feature | Vapi | Retell | Bland | Custom | |---------|------|--------|-------|--------| | Time to first call | Hours | Hours | Minutes | Weeks-months | | Avg. latency | 1,000–1,400ms | 700–1,000ms | 1,000–1,400ms | 600–1,000ms | | Cost per 3-min call | ~$0.16–$0.23 | ~$0.15–$0.24 | ~$0.27 | ~$0.04–$0.06 | | HIPAA | No | Enterprise only | No | Yes (your config) | | LLM choice | Any | Most | Limited | Any | | TTS voice choice | Wide | Wide | Limited | Any | | Custom telephony | Yes (SIP) | Yes (SIP) | Limited | Full | | Complex logic | Moderate | Moderate | Limited | Full | | Outbound dialing | Basic | Basic | Strong | Full (build it) | | Documentation | Excellent | Good | Good | N/A | | Vendor lock-in | Medium | Medium | High | None |
Decision framework
Start with Vapi if: You're validating a concept, you have less than 3 months to launch, your use case doesn't involve PHI, and call volume will be under 10,000/month for at least 6 months.
Choose Retell over Vapi if: Latency matters for your use case (customer-facing sales/support conversations, not automated reminders), or you need HIPAA compliance at the enterprise tier.
Use Bland if: You're running high-volume outbound campaigns with scripted flows, speed of deployment is paramount, and you don't need LLM flexibility.
Build custom if: Any of the following are true — you need full HIPAA compliance with your own infrastructure, call volume exceeds 10,000/month and you want to optimize costs, your conversation logic is too complex for webhook-based integrations, or you're building a product on top of voice AI (in which case platform dependency is a real risk).
Migration paths
One thing to consider: migrating off a hosted platform is painful. Your conversation logic is often encoded in their proprietary config format, your phone numbers are provisioned through them, and your team has built workflows around their tooling.
If there's any chance you'll outgrow a platform, design your system so the business logic lives in your own API server and the platform is just the audio layer. This makes migration dramatically easier — you rebuild the WebSocket pipeline but keep all the conversation logic.
Vapi and Retell both support this pattern via server-side conversation control. Build for it from day one, even if you start on a hosted platform.
The bottom line
For most teams in 2026, the right answer is Vapi or Retell to start — they get you to market fast and handle the infrastructure complexity that's genuinely hard to build. Retell wins on latency and HIPAA; Vapi wins on developer experience and flexibility.
Custom builds are justified when HIPAA compliance, cost at scale, or business logic complexity make the hosted platforms unsuitable. That threshold is lower than most people expect — if you're handling PHI or projecting 10,000+ monthly calls within the next year, custom is worth the upfront investment.
Bland is a narrow-use-case tool: high-volume, simple, outbound. If that's you, it's great. Otherwise, Vapi or Retell will serve you better.