Retell AI Voice Infra Pricing vs Deepgram STT: Are You Overpaying?

[Visual: a split-screen waveform. Left side, a single clean line labeled "Retell: one invoice." Right side, four jagged lines in different colors labeled STT, TTS, LLM, telephony, each billed separately.]

A voice agent call is really four products stapled together: it listens, it thinks, it speaks, it dials. Retell AI sells you all four under one line item called "voice infra." The question worth asking before you sign an annual contract or start building your own stack is simple: when you compare retell ai voice infra pricing vs deepgram stt directly, how much of that bundled fee is actually transcription, and how much is everything else?

This isn't a pricing takedown. It's an accounting exercise. Once you separate the STT line from the orchestration line, the decision about whether to build or buy gets a lot clearer.

What Is Actually Inside Retell AI's Voice Infra Line?

Retell's published pricing breaks a call into four components: a voice infra fee, a text-to-speech fee that varies by provider, an LLM fee that varies by model choice, and a telephony fee. Speech-to-text has no line of its own. It's folded into the voice infra charge.

The published component table

That structure is worth sitting with. Four named costs, and the one everyone assumes is the expensive part, transcription, doesn't appear as its own number anywhere on the page.

Where STT disappears into the bundle

This is a packaging decision, not a technical limitation. Retell runs speech-to-text as part of its own orchestration layer, alongside turn-taking logic, interruption handling, and latency tuning, and chooses to present all of it as one number. The practitioner question that raises: is the voice infra fee mostly margin sitting on top of STT, or mostly the cost of the orchestration and latency engineering wrapped around it?

A bundled price is a design choice about what the customer is allowed to see. It says: don't worry about the parts, trust the whole.

What Does Deepgram Actually Charge for the Same Transcription?

Deepgram prices its Nova-3 model differently depending on how you use it, with separate rates for batch transcription of pre-recorded audio and for real-time streaming. The gap between those two pricing modes matters more than it looks.

Nova-3 batch vs streaming rates

Batch pricing is cheaper because the model can process audio at its own pace, with no latency constraint. Streaming pricing costs more because the model has to produce partial transcripts in near real time, word by word, while the caller is still talking.

Why voice agents need streaming, not batch

Voice agents, the exact use case Retell is built for, require streaming transcription. There's no scenario where a live phone call gets transcribed in batch after the fact and still supports a responsive agent. So the honest comparison point is the streaming rate, not the cheaper batch rate that some napkin-math comparisons reach for.

Deepgram's pricing is also usage-based with nothing else bundled in. No orchestration, no TTS, no LLM routing. It is STT and only STT, scored 8.3/10 by the TopReviewed AI panel, which makes it a clean baseline for isolating what transcription alone actually costs a team assembling its own stack.

[Visual: a single thin price tag pinned next to a much larger price tag, both resting on the same stretch of audio waveform.]

How Much of Retell's Voice Infra Fee Is STT, Really?

Deepgram's streaming STT is priced as a lean, standalone commodity rate, while Retell's voice infra bundle folds STT together with orchestration into a single, noticeably higher per-minute fee. The gap suggests the implied STT cost sitting inside that bundle is a small fraction of the total. The rest is something else.

Doing the subtraction

Even generous padding for turn-taking logic, interruption handling, latency optimization, and failover handling leaves a wide gap between standalone streaming STT pricing and Retell's bundled rate. There's no reasonable accounting method that closes that gap through STT costs alone.

The margin that's left over

That doesn't mean Retell is overcharging. Orchestration is real engineering work: deciding when a caller has finished a thought, handling barge-in when someone interrupts the agent mid-sentence, recovering gracefully when a model call times out. What it does mean is that any marketing framing suggesting "free STT" undersells how much of that fee is non-STT infrastructure dressed up as a transcription line.

The honest framing, stated plainly: Retell isn't hiding an expensive STT markup inside its pricing. It's bundling a cheap commodity, streaming transcription, with a layer of real-time coordination work that Deepgram was never built to do and doesn't attempt to do.

What Does a High-Volume Month Actually Cost on Each Stack?

At meaningful monthly call volume, Retell's bundle totals a mix of voice infra, TTS costs that vary widely by provider tier, LLM costs that vary even more widely by model choice, and telephony. A DIY stack built around Deepgram carries a modest cost for streaming STT, plus separate line items for TTS, LLM calls, and telephony, plus the engineering time to build the orchestration layer Retell ships by default.

Retell's all-in bundle

The width of the LLM cost range for the same volume is the real headline here. Model choice swings the total cost of a Retell deployment far more than the STT question ever could.

Assembling the same stack with Deepgram

Building the equivalent stack independently means sourcing a TTS provider like Eleven Labs, wiring up direct LLM API calls, contracting a telephony provider such as Twilio, and then building the orchestration glue, turn-taking, interruption detection, failover, that Retell provides as a single API call.

Cost component	Retell (bundled)	DIY with Deepgram
Speech-to-text	Included in voice infra	Modest cost (streaming Nova-3)
Orchestration / turn-taking	Included in voice infra	Engineering time, no published rate
Text-to-speech	Varies by provider tier	Priced separately by provider chosen
LLM calls	Varies widely by model choice	Direct API billing, same range applies
Telephony	Modest per-minute fee	Priced separately by provider chosen
Maintenance / on-call	Absorbed by vendor	Ongoing engineering cost

The gap in raw per-minute infra cost is real and easy to point to. What it leaves out is the engineering time required to replicate turn-taking, interruption handling, and failover, which Retell ships without anyone having to design it from scratch.

When Does Unbundling Actually Save Money?

Unbundling saves money fastest for teams that already run Deepgram for something else. If a company already uses Deepgram for support call transcription, analytics, or compliance recording, adding streaming STT to a new voice agent build carries close to zero marginal integration cost.

Teams that already run Deepgram

Picture a 40-person support operations team that already transcribes every inbound call for QA scoring. The Deepgram relationship, the API keys, the billing account, all of it already exists. Standing up a voice agent on the same pipeline is a much smaller lift than starting from zero.

Volume thresholds where DIY pays off

At high volume, the per-minute gap between standalone streaming STT pricing and the implied STT portion of Retell's fee compounds fast enough that building custom orchestration starts to look worth the engineering hours. Teams already running Cloudflare for edge routing or MongoDB for call-state storage tend to be the same teams with the infrastructure maturity to assemble this economically, because half the plumbing is already in place.

The unbundling case weakens fast for teams without an existing real-time infrastructure team on staff. The hidden cost in that scenario was never STT. It's the headcount needed to replace what Retell already built and already operates.

When Does Retell's Bundling Win Instead?

Retell's bundling wins when integration speed is the product being purchased, not cheap transcription. One API call handling STT, turn-taking, interruption detection, and TTS replaces a stack of glue code that would otherwise need a team to build and keep running.

Integration speed as the real product

For teams shipping a voice agent in weeks instead of quarters, the all-in-one rate functions as buying engineering time back, not as overpaying for a transcription API that happens to be marked up.

What you're actually paying for

Observability tools like Honeycomb, scored 8.5/10 by the TopReviewed AI panel, or Sentry become necessary additions the moment a team DIYs a voice stack, because someone now has to monitor latency spikes, dropped streams, and failed turn-taking in production. Retell absorbs a meaningful share of that operational surface internally, which is easy to miss when comparing per-minute numbers in isolation.

The question worth asking isn't whether STT is overpriced inside Retell. It's whether the orchestration premium is worth not hiring for it.

[Visual: two org charts side by side. Left: one engineer, one Retell dashboard tab open. Right: four engineers, five dashboard tabs open, a pager icon blinking.]

How Should You Decide Between the Two Approaches?

Decide by auditing what you already have running, not by comparing per-minute rates in a spreadsheet. If Deepgram, Twilio, and an LLM pipeline are already wired together on your team, unbundling likely saves money at any meaningful volume.

A simple decision framework

• Do you already have Deepgram, a telephony provider, and an LLM pipeline connected for another product? If yes, lean DIY.
• Are you starting from zero with a deadline measured in weeks? If yes, lean Retell.
• Have you modeled your actual LLM cost range, not just the STT question, since that spread often dwarfs everything else in the stack?

Questions to ask before switching

Model your own LLM cost carefully before committing to either path. The range in Retell's own pricing by model choice is the widest variable in the entire stack, and it frequently matters more to the final bill than the STT question that started this comparison.

Testing frameworks like Promptfoo, scored 8.5/10 by the TopReviewed AI panel, are worth running against both stacks before committing to either, since latency and turn-taking quality tend to affect call completion rates more than any single per-minute line item.

Before signing anything, run one week of real call volume through both setups and measure call completion rate, not cost per minute. The stack that finishes more conversations without a dropped turn is the one actually worth paying for.