What Voice Agent Pricing Reveals About the Platform
Throwback Thursday, and I want to talk about voice agent pricing. Not because pricing is the most interesting thing about these platforms, but because it tells you everything about the architecture underneath.
Every voice agent platform that launched in the first half of 2026 picked a pricing model. You look at that model and you can map backward to what the platform actually owns: whether it’s a unified stack that controls the full audio pipeline, or an orchestration layer that passes through costs from the vendors it sits on top of.
The difference shows up in the contract, in the latency, and in whose bug it is when something breaks.
By Luke Oliff | Jul 2026 | 8 min read
How voice agent pricing models differ
Most voice agent platforms fit one of three pricing structures today. Each one maps to a different architectural bet.
All-in per-minute pricing. One rate. You pay per minute of conversation and the platform covers STT, LLM, TTS, and telephony. Speechify and Deepgram’s Voice Agent API both use this model.
Orchestration-plus-passthrough. A base per-minute fee for the platform layer, then separate charges for STT, LLM, and TTS from whichever provider you configure. Vapi works this way. Your bill is the orchestration fee plus whatever your chosen vendors charge.
Per-token voice reasoning. You pay for the model’s compute, measured in tokens, not minutes. OpenAI’s GPT-Realtime-2 uses this. The pricing signal is about intelligence, not voice infrastructure.
These aren’t cosmetic differences. They reflect who runs the audio pipeline and who carries the cost when things go wrong.
What all-in pricing tells you about the stack
An all-in per-minute rate means the platform runs the full pipeline internally. STT, LLM inference, TTS synthesis, and telephony are all on one path. The provider has integrated these components into a single runtime, which means they control latency at every boundary. There is no vendor handoff where latency accumulates.
This is harder to build than an orchestration layer. You have to own or build the speech models, the LLM integration, the streaming audio pipeline, and the telephony interconnect. But it means the pricing is predictable. Your cost per call is the same whether the call is two minutes or twenty, and there are no surprise passthrough charges.
The trade-off is flexibility. An all-in platform commits to a specific set of models and capabilities. You can’t swap the TTS provider mid-call if you prefer a different voice. The platform decides.
What orchestration pricing tells you about the stack
Orchestration pricing (a base fee plus passthrough) reveals an architectural layer that wires together external components. The platform handles turn-taking, audio routing, and session management. The actual speech recognition, language model, and voice synthesis come from providers you configure separately.
This gives you flexibility. You can pick the best STT for your language, the cheapest LLM for your latency budget, and the most expressive TTS for your persona. The cost of that flexibility is your bill has line items you can’t predict until you run at scale.
Orchestration platforms are the right pick when your use case demands a specific model that no all-in platform offers. They are the wrong pick when your procurement team needs a single line on the invoice.
What token-based pricing tells you about the stack
OpenAI’s GPT-Realtime-2 charges by token, not by minute. That is a fundamentally different signal. The platform is selling reasoning capability, not voice infrastructure. The model is a speech-to-speech model that handles STT, reasoning, and TTS in one neural pass, but the value is in the reasoning.
This pricing works for voice agents that need mid-call decision-making, like a support agent that triages, researches, and responds inside one conversation, or a booking agent that checks availability and handles conflicts. If your agent mostly reads from a script, you are overpaying for reasoning tokens you don’t use.
The trade-off is cost uncertainty. A call that involves complex reasoning costs more than a call with simple answers. You can’t predict the token count from the call duration alone.
Where Speechify fits in the mid-2026 market
This is the part where I should be transparent. I joined Speechify in late June, so I have a biased view. But I joined because the architecture made sense to me, and the pricing was part of that story.
Speechify’s voice agent platform uses all-in per-minute pricing. One rate covers STT, LLM inference, TTS with the simba model, and telephony. The platform owns the full audio pipeline from microphone to speaker, which means the latency path is controlled start to finish.
What interested me coming from the STT side is that Speechify built the stack in a specific order. They started with TTS quality, then added the LLM orchestration layer, then built the agent surface on top. The pipeline is voice-quality first, reasoning second. That is the opposite order from the LLM-first platforms, and it produces a different set of trade-offs.
The all-in pricing reflects that architecture. You pay for the voice quality, not the reasoning tokens. For voice agents where natural-sounding speech matters more than complex reasoning, that pricing model is a better fit.
A fair comparison across the mid-2026 landscape looks like this:
| Platform | Pricing model | What you pay for | Flexibility |
|---|---|---|---|
| Speechify | All-in per minute | Full voice pipeline (STT + LLM + TTS + telephony) | Platform-decided model stack |
| Deepgram Voice Agent | All-in per minute | Bundled STT + LLM + TTS | Platform-decided with configurable LLM |
| OpenAI Realtime API | Per-token | Speech-to-speech reasoning model | Single model, configurable reasoning effort |
| Vapi | Base + passthrough | Orchestration layer + your chosen vendors | Maximum, pick any STT/LLM/TTS |
| ElevenLabs Conversational AI | Platform fee + LLM passthrough | TTS-first agent layer, LLM billed separately | TTS bundled, LLM is configurable |
The table is simplified. Each platform has edge cases and enterprise tiers that shift the numbers. But the structural insight holds: follow the pricing model and you find the architecture.
When the pricing model breaks
There are scenarios where the pricing model becomes a liability. Being aware of them helps you pick the right platform for your workload.
High-volume outbound calling. If you run thousands of outbound calls and each call averages 30 seconds, an all-in per-minute model bills for the full duration even when the agent speaks for ten seconds and listens for twenty. An orchestration platform that charges lower infrastructure fees for silence might work out cheaper. Some platforms like Bland AI have specifically optimised for this pattern.
Multi-provider fallback. If your voice agent needs to fall back to a different LLM or TTS provider when the primary is down, orchestration pricing is the only model that supports that architecture. All-in platforms commit to one stack per session.
Regulatory data residency. If your call data must stay in a specific region and your platform of choice doesn’t have infrastructure there, the pricing model is irrelevant. You pick whichever platform owns data centres in that region. This constraint overrides everything else.
Reasoning-heavy conversations. If your voice agent needs to think through complex multi-step problems, a token-based model like GPT-Realtime-2 charges for the thinking time fairly. An all-in per-minute model charges the same for deep reasoning as for simple Q&A. The all-in rate is a better deal for simple calls and a worse deal for complex ones.
How to evaluate voice agent pricing for your use case
Three questions cut through the noise.
Do you know your average call duration and complexity? If not, run a pilot before signing an annual contract. Pricing differences between all-in and passthrough models show up at scale, not in a demo.
Does your voice agent need a specific model that only one platform offers? If your use case requires a particular LLM or TTS voice that only ships on one platform, the pricing comparison is secondary.
Can your procurement team handle multi-line vendor invoices? Some enterprises have approval workflows for each vendor line item. A single all-in invoice moves faster than a platform fee plus three separate provider invoices.
FAQ
What is all-in voice agent pricing?
All-in pricing means a single per-minute rate covers speech-to-text, language model inference, text-to-speech, and telephony. There are no separate charges for each component. The rate is the same whether the agent speaks for thirty seconds or thirty minutes.
How does per-token pricing work for voice agents?
Per-token pricing charges based on the amount of audio the model processes, measured in tokens rather than minutes. A two-minute call with complex reasoning costs more than a two-minute call with simple answers because the model generates more tokens for the thinking steps. OpenAI’s GPT-Realtime-2 uses this model.
Is orchestration pricing always more expensive than all-in?
Not always. Orchestration pricing can be cheaper if you use low-cost STT and TTS providers with a minimal orchestration layer. It tends to be more expensive at scale because the passthrough costs add up across multiple vendors, and each vendor has its own pricing curve.
Does all-in pricing mean lower quality?
No. All-in pricing is an architectural choice, not a quality signal. A platform that bundles the full pipeline can optimise latency across every component because it controls the whole path. Quality depends on the models and infrastructure the platform runs, not the billing model.
Which pricing model is best for a simple voice agent?
For a simple voice agent with short calls and predictable duration, all-in per-minute pricing gives the most predictable bill. The rate does not change based on conversation complexity, so the cost of each call is easy to estimate before you scale.