Last updated: July 8, 2026
A production voice agent built from raw APIs costs roughly $0.007 to $0.091 per conversation minute at July 2026 prices, depending on which STT, LLM, and TTS you pick. This page is a worked cost model from Inworld AI, a research lab and inference provider focused on realtime AI models for consumer-facing applications, priced entirely from public sources so you can rebuild it with your own numbers.
We build modular realtime voice AI models and APIs (Realtime TTS, Realtime STT, the cascaded Realtime API, Inworld-served optimized open models like Gemma 4, and a Router that routes to 220+ models at cost) for high-volume conversational products. One of the six stacks below is ours, priced from the same public pricing pages as everyone else's.
How is voice agent cost per minute calculated?
Voice agent cost per minute is the sum of three metered components: STT (billed per audio minute or hour), the LLM (billed per input and output token), and TTS (billed per character or per audio token). Convert each to a per-conversation-minute figure using explicit usage assumptions, then add them. The assumptions matter more than the vendor logos. Here are ours; swap any of them.
Two honest caveats before the arithmetic. First, the 1.8x history multiplier is a calculator convention, not physics; a 30-minute support call with full history retention will exceed it, and aggressive context truncation will beat it. Second, the 50/50 speaking split favors no one in particular, but a narration-heavy agent (more agent speech) inflates TTS cost and a listening-heavy agent inflates STT share. Measure your own transcripts after week one.
What does each component cost in July 2026?
Per conversation minute, the components below run from $0.0001 (Llama 3.1 8B on Groq) to roughly $0.0912 (OpenAI gpt-realtime-2.1). Every unit price was fetched from the vendor's public pricing page in July 2026, and converted using the assumptions above; third-party or derived figures are labeled. If you are reading this later, re-check the linked pages. These numbers move.
Notes on the two rows that need them. Deepgram's page as fetched on 2026-07-07 lists streaming Nova-3 English at $0.0048/min and pre-recorded at $0.0077/min; we use the streaming rate exactly as published. OpenAI does not publish a per-minute equivalent for gpt-realtime-2.1 on the fetched pricing page, so we convert using 1,500 audio tokens per minute, which is Google's own published equivalence for its Live API ($12.00 per 1M audio output tokens dual-quoted as $0.018/min on
ai.google.dev, fetched 2026-07-07), applied cross-provider as a labeled assumption.
How much does each voice agent stack cost per minute?
Under this model's assumptions, the six stacks land between $0.0069 and $0.0912 per conversation minute at July 2026 prices. The lowest per-minute figure in the table is the fully cascaded Inworld Realtime STT plus Inworld-served Gemma 4 26B plus Inworld Realtime TTS-2 stack, at about $0.007 using Inworld's Growth committed tier (about $0.013 on demand); the highest is OpenAI's flagship speech-to-speech model, at roughly $0.091.
Here is the Inworld row fully worked, so you can rebuild any row the same way. STT: $0.10/hr divided by 60 is $0.001667 per minute at the Growth tier ($0.15/hr on demand;
inworld.ai/pricing, 2026-07-08). LLM: 2,340 input tokens x $0.07 per 1M is $0.0001638, plus 100 output tokens x $0.34 per 1M is $0.000034, so $0.000198 total on Inworld-served, optimized Gemma 4 26B A4B, billed at cost and flat across tiers (
inworld.ai/models, 2026-07-08; the larger Gemma 4 31B lists $0.13 in / $0.38 out). TTS: 400 characters x $12.50 per 1M characters is $0.0050 at the Growth tier ($25 per 1M on demand;
inworld.ai/pricing, 2026-07-08). Sum: about $0.00686 per conversation minute, or $6.86 per 1,000 minutes.
The speech-to-speech rows carry the widest error bars. The audio-token-per-minute conversion is our labeled assumption for OpenAI, and history handling in realtime sessions (whether prior agent audio re-bills as input) can push the flagship figure meaningfully higher on long calls. The Gemini Live row is the most trustworthy speech-to-speech number here because Google publishes the per-minute rates directly.
At consumer scale the spread compounds: 100,000 conversation minutes per month is about $686 on the Inworld Growth-tier stack versus about $9,120 on gpt-realtime-2.1, at the same prices and assumptions. If you are still choosing an architecture before choosing vendors, start with
how to build an AI voice agent and the
STT to LLM to TTS pipeline fundamentals.
Get an API key at
platform.inworld.ai and price your own stack against real traffic; the free tier includes up to 70 minutes of TTS, so your first measured cost-per-minute number costs nothing.
What dominates voice agent cost: TTS characters or LLM tokens?
TTS dominates every cascaded stack in this model: 70 to 73 percent of raw component cost at 400 characters per conversation minute. The LLM is under 4% of cost on either small open model above. Optimize TTS pricing first; token pinching is a rounding error until your model choice changes.
The ranking flips only at frontier LLM rates. Swap the Inworld-served Gemma 4 26B for OpenAI's flagship gpt-5.5 ($5.00 input, $30.00 output per 1M tokens,
OpenAI pricing, fetched 2026-07-07) and the LLM line alone becomes 2,340 x $5/1M plus 100 x $30/1M, about $0.0147 per minute. That single substitution costs more than the entire Inworld cascaded stack, which is the practical argument for routing most turns to a small model and escalating selectively.
Committed volume moves TTS just as sharply. Inworld Realtime TTS-2 is $12.50 per 1M characters at the Growth tier used above (down from $25 on demand) and drops as low as $5 at enterprise volume (
inworld.ai/pricing, 2026-07-08), taking the Inworld stack from about $0.0127 on demand to $0.0069 at Growth to about $0.0039 at enterprise TTS rates per minute. Deepgram's Growth tier prices Aura-2 at $0.027 per 1K characters versus $0.030 pay-as-you-go (
deepgram.com/pricing, 2026-07-07). ElevenLabs' subscription tiers work out to roughly $100 per 1M characters of included quota at the Multilingual rate, and about half that consuming credits on Flash (derived from tier price divided by included characters,
elevenlabs.io/pricing/api, 2026-07-07).
What hidden costs change the math?
List prices are the floor, not the bill. The items below are the recurring gap between a model like this one and the invoice, and none of them appear in a per-character or per-token rate card.
- Minimum commitments and subscription floors. Deepgram's Growth rates require $4K+ per year prepaid (deepgram.com/pricing, 2026-07-07). ElevenLabs' effective rates depend on monthly tiers from $6 to $990 with included credit quotas (elevenlabs.io/pricing/api, 2026-07-07); unused quota is spend without minutes.
- Session-billed versus audio-billed STT. AssemblyAI bills streaming per WebSocket session time, not audio duration (assemblyai.com/pricing, 2026-07-07). An agent that holds sockets open through silence pays for the silence. This model's 1.0 STT-minute assumption reflects session billing; audio-submitted billing can cut the STT line roughly in half at a 50/50 speaking split.
- Platform per-minute markup versus raw APIs. Managed platforms bill wall-clock minutes at rates that bundle orchestration: ElevenLabs Agents at $0.08/min, Cartesia Line at $0.06/min, Deepgram Voice Agent API at $0.050 to $0.163/min (all fetched 2026-07-07 from the vendors' pricing pages). Against a $0.007 to $0.029 raw stack, that is roughly a 2x to 24x premium for not running the pipeline.
- Telephony. A Cartesia-provided phone number adds $0.014/min (cartesia.ai/pricing, 2026-07-07). A DIY stack needs its own SIP trunking, which appears in none of the rate cards above.
- History growth on long calls. The 1.8x multiplier here models short conversations. Retained context grows input tokens roughly linearly per turn, so 20-minute calls with no truncation can double or triple the LLM line.
- Barge-in waste. Every interruption discards synthesized-but-unplayed TTS audio you already paid for. Agents with poor turn-taking regenerate 10 to 30 percent extra characters; that multiplier applies to your largest cost line.
- STT add-ons. Extras stack on base rates: AssemblyAI lists speaker ID at +$0.02/hr, entity detection at +$0.08/hr, and sentiment at +$0.02/hr (assemblyai.com/pricing, 2026-07-07).
When is another stack or a platform the better choice?
No single stack wins every deployment, and raw component list prices are not the whole cost picture. The honest decision boundaries at July 2026 prices look like this.
A managed platform beats raw APIs when your volume is low and engineering time is the scarce resource. At 10,000 minutes per month, the difference between Cartesia Line at $0.06/min ($600) and a DIY Inworld stack at $0.0069/min ($69) is $531, which does not fund the orchestration, telephony, and on-call work the platform absorbs. Compare options in our
voice agent platform comparison and, if you are weighing frameworks instead, the
Vapi vs Pipecat vs LiveKit breakdown covers framework platform fees.
OpenAI Realtime is worth its per-minute price when you want one vendor end to end and native audio-in, audio-out expressivity, such as prosody-aware responses that a cascaded pipeline approximates less directly. gpt-realtime-2.1-mini at roughly $0.029 per minute under these assumptions lands in the same per-minute range as the ElevenLabs plus OpenAI cascaded stack.
ElevenLabs is the better TTS choice when your product is content-first rather than conversation-first: its Multilingual v3 model targets 70+ languages with expressive audio tags (
elevenlabs.io, fetched 2026-07-07), and for audiobooks or dubbing the per-character premium buys a large voice library. Realtime TTS-2 (research preview) supports 15 production languages, expanding to over 100 with experimental support.
The budget Deepgram plus Groq stack is the right call for English-only, latency-tolerant, cost-floor deployments where voice quality is secondary, such as internal tools or high-volume notifications. And phone-first deployments have their own constraints; see
best voice AI for phone agents for the telephony-specific comparison.
About Inworld AI
Inworld is a research lab and inference provider focused on realtime AI models for consumer-facing applications. We build first-party voice models (Realtime TTS and Realtime STT), serve optimized open-source LLMs on our own Realtime Inference engine, and expose them as modular APIs, alongside an LLM Router that routes to 220+ models and a Realtime API for full speech-to-text-to-LLM-to-speech pipelines. We focus on serving developers of realtime, high-volume conversational products across domains such as health, fitness, education, companions, social, and games, with an emphasis on quality, low latency, and low cost at scale.
Frequently asked questions
How much does a voice agent cost per minute in 2026?
At July 2026 prices, a voice agent built from raw APIs costs roughly $0.007 to $0.091 per conversation minute depending on the stack. A cascaded pipeline of Inworld Realtime STT, Inworld-served Gemma 4 26B, and Inworld Realtime TTS-2 computes to about $0.007 per minute at Inworld's Growth committed tier (about $0.013 on demand). An ElevenLabs plus OpenAI cascaded stack computes to about $0.029 per minute. OpenAI's gpt-realtime-2.1 speech-to-speech model computes to roughly $0.09 per minute under the same usage assumptions. Managed voice-agent platforms charge $0.05 to $0.163 per minute on top of, or instead of, raw component pricing.
What is the biggest cost driver in a voice agent?
Text-to-speech, by a wide margin, in every cascaded stack we priced. At about 400 TTS characters per conversation minute, TTS is roughly 70 to 73 percent of raw component cost, STT is second, and the LLM is under 5 percent if you use a small open model. The ranking flips only if you use a frontier LLM: OpenAI's gpt-5.5 at $5 input and $30 output per 1M tokens adds about $0.015 per minute in LLM cost alone, more than the entire Inworld cascaded stack.
How does OpenAI's Realtime API compare in cost to a cascaded STT, LLM, TTS pipeline?
At July 2026 prices, the arithmetic favors cascaded pipelines. Under identical usage assumptions, gpt-realtime-2.1 ($32 per 1M audio input tokens, $64 per 1M audio output tokens) computes to roughly $0.091 per conversation minute, versus $0.007 to $0.029 for the cascaded stacks in this model. gpt-realtime-2.1-mini ($10 and $20 per 1M audio tokens) lands near $0.029 per minute, in the same range as the ElevenLabs plus OpenAI cascaded stack. Speech-to-speech buys you a single vendor and native audio expressivity; the per-minute arithmetic is the price of that simplicity.
Why do voice agent platforms charge more per minute than raw APIs?
Because a per-minute platform price bundles orchestration, session infrastructure, and often telephony that you would otherwise build. As of July 7, 2026: ElevenLabs Agents lists $0.08 per minute, Cartesia Line lists $0.06 per minute plus $0.014 per minute for a Cartesia-provided phone number, and Deepgram's Voice Agent API lists $0.050 to $0.163 per minute depending on tier and BYO configuration. That is roughly 2x to 24x the raw component cost of the cascaded stacks in this model, which is the price of not running the pipeline yourself.
How many TTS characters does a voice agent use per minute?
Budget about 800 characters per minute of agent speech. Typical conversational English runs about 150 words per minute, and vendor and analyst estimates put English narration at roughly 700 to 900 characters per minute (
Camb.ai's TTS price comparison, fetched July 7, 2026). If the agent speaks about half of each conversation minute, that is about 400 billable TTS characters per wall-clock minute. Multiply your provider's per-character rate by your own measured character volume; production transcripts beat any published assumption.
How can I reduce voice agent cost per minute?
Attack TTS first, since it dominates cascaded cost. Committed-use tiers move the number materially: Inworld Realtime TTS-2 goes from $25 per 1M characters on demand to $12.50 at the Growth tier and as low as $5 at enterprise volume (
inworld.ai/pricing, July 8, 2026), which takes the Inworld stack in this model from about $0.0127 on demand to $0.0069 at the Growth tier used here to about $0.0039 at enterprise volume. Then cap LLM history growth (the 1.8x multiplier here), use a small open model for most turns, and check whether your STT bills socket time or audio submitted.
Next steps