Vapi vs Retell vs Bland: The Real Cost Per Minute in 2026

There is no single cheapest platform - the headline rates are not comparable, and once you price the full stack, real costs converge to $0.10-$0.30 per minute across all three. Vapi is the most flexible and, with careful component selection, can be the cheapest, but only if you are willing to manage separate LLM, STT, and TTS accounts. Bland is the simplest to budget because almost everything is bundled into one rate. Retell sits in the middle on both flexibility and predictability.

The right choice depends less on the per-minute number and more on how much engineering time you want to spend optimizing voice-stack components versus shipping the agent. Teams optimizing hard for cost at scale (tens of thousands of minutes a month) benefit from Vapi's component control. Teams that want a working phone agent this week without three separate vendor integrations lean toward Bland or Retell. As of July 2026 - verify current pricing before committing, since all three vendors adjust rates and bundles frequently.

If you are earlier in the decision - not sure a paid voice platform is even the right starting point - our no-code voice AI agent guide is a gentler on-ramp before you commit to per-minute billing at all.

This is the table that actually matters, because comparing headline rates without knowing what each one includes is comparing apples to invoices.

The pattern: Vapi's low headline number is real but incomplete - it is a platform fee, not a call cost. Speech-to-text typically runs $0.005-$0.02/min and the LLM turn cost another $0.02-$0.10/min depending on model choice, so the "$0.05/min" platform quickly becomes $0.10+ once STT, LLM, and TTS are added. Bland's flat rate is the most honest headline number because there is little left to add. Retell is the middle ground - telephony is bundled so you have fewer separate bills than Vapi, but LLM/TTS choice still swings the final number across a wide range.

Two practical consequences follow from this table. First, sales conversations and marketing pages that quote a single per-minute number without specifying what it includes should be treated as a starting point, not a quote - ask directly which of STT, LLM, TTS, and telephony are bundled versus billed separately, and get the answer in writing before sizing a budget. Second, the "cheapest" platform on paper can become the most expensive in production if the BYOK components are chosen carelessly - a team that defaults to a premium TTS voice and a large frontier LLM on Vapi can easily end up paying more per minute than Bland's flat rate, even though Vapi's platform fee alone is the lowest of the three.

Take a modest but real deployment: a support or scheduling line handling 1,000 minutes of AI-agent conversation per month (roughly 250 four-minute calls). Here is the all-in monthly cost on each platform using mid-range component choices - a mainstream STT provider, a cost-efficient LLM (not the largest frontier model), and a standard-tier TTS voice.

At 1,000 minutes, the gap between platforms is real but not dramatic - roughly $20-$85/month depending on component choices, not the 2-3x the headline rates alone would suggest. The gap widens in either direction at scale: at 50,000+ minutes/month, Vapi's component control lets a team that optimizes hard (cheaper STT, a smaller LLM for simple turns, a budget TTS tier) undercut both competitors meaningfully; a team that does not optimize and just takes default components can end up paying more on Vapi than on Bland's flat rate. These figures are illustrative estimates built from published component pricing patterns, not vendor quotes - as of July 2026 - verify current pricing and run your own component costs before budgeting a production deployment.

For voice agents, latency is as much a product-quality issue as a cost one - callers notice a slow response far more than a chat user does. Latency is driven mostly by the STT-LLM-TTS pipeline architecture, not just the orchestration layer, so it correlates more with which components you choose than which platform you use.

• Vapi - because you choose every component, latency is the most controllable: pairing a fast STT (streaming, low-latency mode), a quick-responding LLM, and a low-latency TTS engine can get end-to-end response times competitive with the fastest bundled options. The tradeoff is that a poorly chosen component stack (a slow frontier LLM, a high-quality-but-slow TTS voice) can also make Vapi the slowest of the three.
• Retell - bundled telephony and a curated set of LLM/TTS options generally deliver consistent, well-tuned latency out of the box, since Retell has pre-optimized the integration between its supported providers.
• Bland - fully bundled and tuned by Bland end to end, latency is consistent and does not require any configuration, but you cannot trade voice quality for speed the way you can on Vapi or Retell.

If sub-second response time is a hard product requirement (e.g., replacing a live phone agent where callers will hang up on dead air), test actual round-trip latency with your specific component choices before committing - published averages vary by model load, region, and time of day, and no platform's marketing number is a reliable substitute for your own test calls.

Voice agents that handle customer calls touch more regulatory surface than most text-based agents: call recording consent laws, HIPAA if health information is discussed, PCI DSS if payment details are collected, and TCPA rules around outbound and automated calling in the US.

• All three platforms offer some combination of SOC 2 compliance, HIPAA-eligible configurations (typically on higher/enterprise tiers with a signed BAA), and call recording/consent controls - but exact certifications and BAA availability change over time and by plan tier, so treat any specific claim here as something to confirm directly with the vendor, not something to take from a comparison article.
• Bring-your-own-key architectures (Vapi) mean compliance is partly your responsibility to assemble correctly - your LLM provider, STT provider, and TTS provider each need their own compliance posture (e.g., a BAA with your model provider, not just with Vapi) if you are in a regulated use case. Bundled platforms (Bland, and Retell for its bundled components) centralize that responsibility but reduce your visibility into subprocessor chains.
• Outbound calling in particular carries TCPA exposure in the US regardless of platform - the compliance burden of consent and do-not-call handling sits with you as the caller of record, not the platform vendor, on all three.

If your use case involves healthcare, finance, or outbound calling to consumers, get compliance details in writing from the vendor's current documentation before building - this is one area where "as of July 2026" guidance ages fast and general advice is not a substitute for a signed agreement.

The right platform tracks the shape of the deployment more than any single spec. Three examples make the tradeoffs concrete.

Scenario 1: a dental clinic piloting an appointment-reminder and rescheduling line. Call volume is modest (a few hundred minutes a month), the conversation flows are simple and repetitive (confirm, reschedule, cancel), and there is no in-house engineering team to manage separate STT/LLM/TTS vendor accounts. Bland's flat, bundled rate is the natural fit here: one invoice, no component decisions, and the simple conversation flows do not need Vapi's flexibility or Retell's model choice to perform well.

Scenario 2: a mid-market SaaS company replacing a tier-1 support phone queue. Volume is meaningful (tens of thousands of minutes a month), call content varies widely (billing questions, technical troubleshooting, escalations), and the team wants some control over which LLM handles which call type without owning the full stack. Retell's hybrid model fits well: bundled telephony removes one integration, while LLM and TTS choice lets the team route simple calls to a cheaper model and complex ones to a stronger one.

Scenario 3: a voice AI startup building a product on top of a voice pipeline, optimizing hard for unit economics at scale. Volume is large (hundreds of thousands of minutes a month), engineering resources exist to manage multiple vendor integrations, and every fraction of a cent per minute compounds across the customer base. This is Vapi's use case: BYOK access to the cheapest viable STT, a right-sized LLM per call complexity tier, and a TTS provider chosen on a cost/quality curve the team controls directly. The integration overhead that would be wasted effort in scenario 1 is the entire point in scenario 3.

The pattern across all three: low volume and low in-house capacity favor bundled platforms (Bland, then Retell); high volume and available engineering time favor component control (Vapi). Most real deployments sit somewhere between scenario 1 and scenario 2, which is why Retell's hybrid model is often the least risky default when you are not sure yet which way your usage will grow.

• You have (or want) engineering time to own the STT/LLM/TTS/telephony stack independently and optimize each component for cost or quality.
• You are running at high volume where component-level optimization meaningfully moves the total bill.
• You want to swap providers quickly - test a new TTS voice, try a cheaper LLM for simple call types - without waiting on a platform vendor to add support.
• Your team already has accounts and expertise with providers like Deepgram, ElevenLabs, and a major LLM API.

• You want to ship a working phone agent quickly without integrating and billing four separate vendors.
• Predictable, easy-to-forecast per-minute cost matters more than squeezing the lowest possible cost per component.
• Your team does not have (or does not want to build) in-house expertise in STT/TTS provider selection and tuning.
• Choose Retell specifically if you still want some control over LLM and voice choice but do not want to manage telephony separately.
• Choose Bland specifically if you want the simplest possible billing - one line item, no component reconciliation - and are comfortable with less granular control over the underlying models and voices.

Per-minute pricing is the most visible cost lever, but it is rarely the only one worth optimizing once a voice agent is in production. LLM token costs scale with conversation length and model choice independent of which voice platform you use, and the same cost-reduction techniques that apply to text agents - shorter system prompts, cheaper models for simple turns, caching where the platform supports it - apply here too. Our guide to reducing AI agent LLM costs covers those techniques in more depth, and they compound with whichever voice platform you choose.

Before comparing per-minute rates at all, it is worth stepping back and running a structured evaluation of whether a given tool - voice platform or otherwise - actually fits your use case, team, and volume. Our framework for evaluating AI agent tools is the same lens applied more generally, and if you want guided, hands-on practice building and shipping a voice agent rather than just comparing vendors, our voice AI agents course walks through the build end to end.