The platform tax: what you're actually paying for a managed voice agent

Managed voice agent platforms advertise low per-minute orchestration rates. When you track every invoice, fully loaded costs can climb once LLM tokens, text-to-speech, and telephony bill through.

The linked Hermes cost analysis places stacks pairing GPT-4o with premium voices near the top, at several times the advertised orchestration rate. Architecture determines which costs appear on the invoice and which ones move into engineering work or visibility gaps.

This article breaks down the three hidden costs behind every voice agent stack and compares how four providers price them. It also gives you the contract questions that surface your real cost before you sign.

Sticker rates and real voice AI pricing diverge in several ways.

• Headline orchestration rates often expand once models, TTS, and telephony bill through.
• Every architecture shifts structural cost into markup, engineering work, visibility gaps, or some combination of them.
• Your production stack can combine two layers and pay two of these costs at once.
• Transparent pricing means published bundled rates and invoices you can decompose at renewal, including lower rows for BYO configurations.

Four providers (Vapi, Retell AI, Bland AI, and Deepgram) publish per-minute rates, but each rate covers a different mix of components, so the numbers aren't directly comparable at face value.

Competitor pricing notes use the public rate categories shown in the table rather than treating every minute as equivalent. Deepgram's published rates appear in the transparent-pricing section. Because vendor pricing changes, verify competitor pricing, pass-through, model-menu, add-on, and concurrency details before signing.

You won't find every telephony, concurrency, or compliance detail in the table. Some pricing pages separate managed telephony and let you avoid that line with your own SIP trunk. Concurrency fees and add-ons behave the same way. They stay off the base rate until the invoice arrives.

Model the full stack before you trust a sticker rate. Most vendors price only one or two services in a five-part stack.

On pass-through platforms, the headline rate can cover orchestration and hosting while the transcriber, model, voice, and telephony bill separately at provider cost, an honest structure, but one that makes the advertised number the smallest line item on your bill. Other managed rates may fold selected components into their own rates while leaving LLM or telephony separate.

You'll pay the LLM separately on both pass-through models, and it's usually the largest line. The total can climb quickly once TTS and an LLM are added. Add-ons stack further: knowledge base retrieval and PII removal add their own per-minute charges.

A bundled all-in rate can look higher than a pass-through headline, yet it may fold LLM, STT, and TTS into the per-minute price. Use your modeled cost per minute for your own call profile.

Pick an architecture and you inherit its cost profile: managed platforms carry a markup, DIY frameworks carry engineering overhead, and speech-to-speech models carry a visibility gap.

When a managed platform prices premium third-party voices above default voices, it creates a spread you can't negotiate directly, and this is the platform tax in its mildest form: convenience priced into components you could buy directly. The orchestration fee itself is part of the same tax, and it's the part you can see.

Assembling the stack yourself makes the per-minute rate fall and the engineering bill rise. Pipecat's self-hosting docs are blunt about why: "A single session ties up a process ... for the entire duration of a conversation."

According to the docs, that kind of work can't be paused, migrated, or interrupted once a call is underway without risking a dropped session. Deploys have to drain live sessions or risk dropping calls. None of that appears per minute; it appears as headcount. Anyone who's carried the on-call pager during a bad deploy already knows this cost by feel.

Audio tokens go in and audio tokens come out, and one line bills for everything between. OpenAI's Realtime API prices audio input and audio output by token, with output audio priced above input audio.

When a call fails or a cost spikes, there's no per-stage boundary to inspect. The transcription, reasoning, and speech a cascaded pipeline itemizes are one model's internal state here. You pay this tax in engineering hours spent reconstructing what happened. If you've ever grepped through logs at midnight trying to explain a cost spike, you've paid it too.

Beyond the markup itself, the same contract structure caps what you can tune and blurs what you can benchmark at renewal.

Concurrency fees show how line items stack past the base rate. If your plan includes a base concurrency allowance, added capacity can bill monthly, a line that won't show up in a per-minute comparison but still sets a floor under your monthly bill regardless of minutes used.

You can only tune what the platform exposes. Model menus are the visible edge of this: the LLM choices you get are the ones listed, at the prices listed. Behavior below the API surface, such as turn-taking, endpointing, and retry logic, belongs to the platform.

When a production edge case needs a fix one layer down, the ceiling is where your options shrink to filing a ticket or migrating. Every platform engineer knows that particular flavor of stuck.

Two years in, benchmarking your rate means decomposing an invoice the platform never itemized. The bundling that simplified procurement now hides which component drives spend, so you can't compare a renewal quote against buying components directly.

Tiered structures deepen this. Plans that trade monthly fees for lower per-minute rates make your effective rate depend on the volume assumptions you locked in at signing.

Production deployments usually combine two architectural layers, which means you may carry two of these costs at the same time without itemizing either.

You own deployment risk while the realtime model owns the visibility gap when Pipecat sits in front of speech-to-speech. You carry the integration tax through session draining, scaling, and telephony plumbing.

Even so, you still can't attribute a failed call or cost spike to a specific stage. The combination looks lean because it avoids a platform fee, yet it stacks the two costs hardest to see on an invoice.

In a managed model menu, a realtime speech-to-speech option can sit far above a smaller text model inside one dropdown. Choose the realtime model and you pay the platform's rate plus the priciest line on the menu.

The voice infrastructure rate and realtime model line can bill at the same time. The observability tax stacks on top, since the model's internal stages stay opaque no matter who bills you.

Every architecture carries a cost, so the useful question is which one you can see, measure, and negotiate. Pick the cost that shows up as an itemized line on an invoice, since that's what you can price against alternatives at renewal. Engineering overhead and observability gaps don't come with a bill you can benchmark.

A pricing page should show included components and how pricing changes when you swap a component or reach ten times your volume.

You shouldn't have to renegotiate blind to get scale economics. Deepgram's pricing page publishes current rates for the Voice Agent API, with Growth-tier discounts listed alongside.

Usage-based billing gives you a published figure to compare against competitor per-minute rates. You don't have to convert token counts into a rate to line the numbers up. For high monthly volume, the Enterprise plan routes pricing through custom commercial terms.

On the Voice Agent API, swapping a component lowers the bill instead of forcing a renegotiation. The Voice Agent API lists BYO LLM and BYO TTS configurations at reduced published rates.

Bringing your own model becomes a pricing row rather than a negotiation. The bundled configuration runs on Deepgram's own speech-to-text stack and text-to-speech models, and either side stays swappable when requirements change.

Bundling solves the five-invoice problem only if the bundle stays decomposable. One contract covering orchestration, transcription, LLM processing, and speech generation removes four vendor relationships without hiding which piece you'd swap. Decomposable bundling gives you voice AI pricing you can audit instead of a bundle that only looks simple.

Model next year's volume before you commit to this year's rate. The contract that looks cheapest at 10,000 minutes a month can be the most expensive one at 100,000.

Ask vendors for written answers before you sign:

1. What's my fully loaded per-minute cost at three times my current volume, including LLM pass-through, telephony, concurrency fees, and add-ons?
2. Which components can I swap or bring myself, and what does my rate become when I do?
3. Can I export recordings and transcripts at exit, and can I take configuration with me?
4. Can you itemize my invoice by component so I can benchmark renewal?

If a vendor can't produce a component-level breakdown, you've found your platform tax. When the answer to swapping is "Enterprise only," that is your abstraction ceiling. To model your real per-minute cost at production volume before signing, run the numbers against a full stack breakdown.

Then compare what you get against the field. The voice AI buyer's guide ranks the major platforms on exactly these dimensions.

For a quick baseline, you can grab free credits and test a representative call with $200 in free credits.

It's the premium between a platform's bundled component prices and what those components cost directly, plus orchestration. Ask procurement for invoice lines such as orchestration, model, voice, transcription, telephony, phone numbers, and concurrency. If the delta stays hidden across default, BYO LLM, and BYO TTS quotes, the tax isn't auditable.

Billing mechanics differ beneath similar-looking rates. A warm transfer might create two telephony legs: one for the caller and one for the destination. Some contracts also add transfer fees, post-call summaries, retained recordings, phone numbers, or monthly concurrency capacity outside the base minute.

Not always. BYO shifts work back to you: authentication, retries, logging, data retention, rate limits, and provider outages. Compare the managed quote against your model bill plus engineering time, incident coverage, and compliance review. BYO wins when control and lower model spend outweigh that operating work.

Cache misses, verbose responses, retries, and failed calls can move spend quickly. Set alerts against your own baseline for repeated retries and calls with unusually long output audio or no successful resolution. Keep prompts, transcripts, and audio metadata separate enough that invoice surprises don't become archaeology.

Use a spreadsheet formula: ((orchestration + STT + LLM + TTS + telephony + add-ons) * minutes + monthly fees) / minutes. Run it for today's traffic and your renewal-year forecast, with a higher-volume scenario such as three-times traffic in between.