Short answer: Mistral’s Voxtral 3, launched 20 October 2026, sets the new floor at $0.04 per minute for a full voice agent with the reasoning model included. OpenAI’s GPT-Live 1 follows at $0.05; Deepgram and ElevenLabs sit at $0.075–$0.08 — but the headline rate is not the bill. What is bundled with it (the LLM, the telephony) moves the real per-call cost more than the brand name does.
Voice AI is having its pricing-war moment: in one October week, Mistral launched Voxtral 3 and ElevenLabs launched Eleven Realtime, joining Deepgram’s Voice Agent API and OpenAI’s Realtime API. For a small business eyeing an AI that answers phones, the question is no longer whether the technology works — it is what a call costs, and what each per-minute figure leaves out.
This is a pricing comparison, not a quality review: both launches are days old, and no independent production benchmark of either exists yet. What we can compare honestly is the money.
Voice agent pricing at a glance
| Platform | Headline agent rate | What the rate includes | Billed on |
|---|---|---|---|
| Mistral Voxtral 3 | $0.04/min (Small 4 reasoning); $0.08/min (larger model) | Transcription + reasoning LLM + speech generation, one price | Per minute of agent use |
| Deepgram Voice Agent API | $0.075/min (Standard) | STT + LLM + TTS bundled (BYO tiers: $0.065/$0.050/$0.163) | WebSocket connection time |
| ElevenLabs Agents | $0.08/min | Speech engine only — LLM and telephony billed separately | Per call minute (burst: $0.16) |
| ElevenLabs Eleven Realtime | $0.08/min connected audio + tokens | Speech-to-speech model; input/output tokens on top | Connected audio minutes + token counts |
| OpenAI Realtime API | GPT-Live 1: $0.05/min; gpt-realtime-2.1: $32/$64 per 1M audio tokens | Model only — backend model/tool usage and telephony are extra | Per-second sessions / audio tokens |
Vendor pricing pages, verified 23 October 2026. Telephony is extra everywhere — about a cent a minute via SIP providers. Eleven Realtime figures via i-scoop’s 21 Oct 2026 read of ElevenLabs’ API page.
Mistral Voxtral 3: the new price floor, with the LLM inside
Voxtral 3 is the reason this comparison exists this week. Mistral’s announcement (20 October 2026) bundles transcription in 50+ languages, text-to-speech in 35+ languages, and a LiveKit-based voice-agent SDK that streams listening, reasoning, tool calling and speaking as one pipeline — the same agentic pattern as OpenAI’s Dots and Meta’s Muse. It answers and places calls over SIP and runs in browsers over WebRTC.
The pricing is the disruption: the full voice-agent API costs $0.04 per minute with Mistral Small 4 reasoning, $0.08 with a larger model — one rate covering transcription, the LLM’s thinking, and the spoken reply. Nobody else here bundles the reasoning LLM into the per-minute price, which is structurally why Voxtral 3 undercuts the field by roughly half on paper.
Two caveats. Every performance claim — the 4.1% FLEURS word error rate, sub-500ms latency — is vendor-reported, with no independent evaluation or named customers yet; clean-benchmark scores routinely degrade on real telephone audio. And the components are sold separately too (batch transcription from $0.0035/min, realtime $0.008/min, TTS $13 per million characters), which matters if you need one pipeline stage, not the full agent.
The other differentiator is deployment: EU, US and global endpoints, plus full self-deployment by arrangement. For a European business with data-residency rules — or a clinic that cannot let call audio leave its environment — none of the other three offers that path.
ElevenLabs: two new products, and the LLM is extra
ElevenLabs answered within a day with Eleven Realtime (v1 Mini and v1 Max), a speech-to-speech model handling listening, turn-taking, reasoning and speaking as one system. The headline rate is $0.08 per minute of connected audio on both — then the tokens: input $0.20/$1.80 and output $6/$9 per million on Mini/Max. ElevenLabs does not publish tokens per minute, so the true cost is an estimate: about $0.084/min (Mini) and $0.102/min (Max) at typical usage, per i-scoop’s 21 Oct 2026 read of the page. Our ElevenLabs review covers the broader platform.
Alongside Realtime sits ElevenLabs Agents at $0.08 per call minute — the speech engine only. The LLM is billed at provider cost on top (ElevenLabs’ own example: Gemini 2.5 Flash at $0.0029/min), and telephony is billed at cost by your provider — ElevenLabs adds nothing for it. Plans bundle minutes: 15 free, then 75 ($6/mo), 275 ($22), 1,238 ($99), 3,738 ($299) and 12,375 ($990). All figures from ElevenLabs’ agents pricing page, verified 23 October 2026.
The honest read: ElevenLabs is the voice-quality leader of the four, but its agent pricing has the most moving parts. Budget the $0.08, add a few tenths of a cent for the LLM, add your phone line — and watch concurrency, because minutes above your limit burst at $0.16.
Deepgram: the all-in bundle for teams that want one invoice
Deepgram’s Voice Agent API (Standard) costs $0.075 per minute pay-as-you-go ($0.068 on Growth), bundling speech-to-text, LLM and text-to-speech — the closest to Mistral’s all-in approach. BYO tiers: $0.065/min with your own TTS, $0.059 with your own LLM, $0.050 with both; Advanced is $0.163/min. Billing is per minute of WebSocket connection time — hold music and silence count. Figures from Deepgram’s pricing page, verified 23 October 2026; new accounts get a $200 free credit, no card required.
Deepgram’s raw components are even cheaper (Nova-3 streaming transcription from about half a cent per minute) — which is why many production agents are cascaded Deepgram + cheap-LLM + TTS stacks rather than the bundled API. One vendor and one meter, or cheapest parts with your own assembly.
OpenAI Realtime API: a $0.05 floor, then tokens
OpenAI’s current voice lineup is two products with two meters. GPT-Live 1 voice sessions are billed at $0.05 per minute — per second, without rounding up — but backend model and tool usage is charged separately on top. The gpt-realtime-2.1 models are token-metered: $32 per million audio input tokens and $64 per million output tokens, with a mini variant at $10/$20. All figures from OpenAI’s API pricing page, verified 23 October 2026. There is no free credit on the page and no phone-number service: telephony is entirely yours to arrange.
The familiar OpenAI pattern: a cheap, flexible model, with SIP trunking, orchestration and concurrency as your engineering. Managed platforms fill the gap — Vapi about $0.05/min on top of your own keys, Retell roughly $0.07–$0.31/min all-in — and at volume that platform fee dwarfs the model-cost differences between vendors.
What one call actually costs: the planner
Billing units differ — connected audio, WebSocket time, tokens — so run your own numbers. The planner totals each platform’s AI fees, adds a telephony line, and shows the per-call cost. Telephony defaults to $0.01/min as an example SIP rate; ElevenLabs Agents adds the LLM at provider cost (ElevenLabs’ example: $0.0029/min).
Voice agent monthly cost planner
How the maths works. Monthly minutes = calls per day × working days × average call length; each platform’s AI fee = minutes × its effective per-minute rate (ElevenLabs Agents includes ElevenLabs’ example LLM rate of $0.0029/min). Add telephony at your $/min on the same minutes; per-call cost = total ÷ calls. Assumption: every minute bills the same, which flatters connection-time billing — Deepgram’s WebSocket-time billing and ElevenLabs’ burst pricing both get worse under spiky traffic. Our calculation; vendor rates as cited above.
On the defaults — 30 calls a day, 22 working days, 4 minutes a call, a cent a minute for the line — Mistral’s $0.04 tier totals about $132 a month ($0.20 a call), OpenAI’s GPT-Live 1 near $158 ($0.24), and the $0.075–$0.08 cluster $224–$248 ($0.34–$0.38). The platform decision is worth about $100 a month here: real money, but less than the cost of building it badly — which is why the next section matters more than the table.
The billing unit matters more than the brand
Four traps hide in these price pages:
1. “Per minute” means four different things. Mistral bills agent minutes. Deepgram bills WebSocket connection time — hold music counts. ElevenLabs Realtime bills connected audio plus tokens. OpenAI bills tokens. A 4-minute call with 90 seconds of hold music costs 4 minutes on one platform and more on another. Your call recordings, not the marketing page, tell you which billing unit fits your traffic.
2. The LLM is the swing cost. Mistral and Deepgram bundle the reasoning model; ElevenLabs and OpenAI do not. A cheap agent minute with an expensive reasoning model behind it can cost more than an expensive minute with a cheap one — which is why Mistral’s $0.04 needs the qualifier “with Mistral Small 4”. Take the larger-model tier and the headline doubles to $0.08 before you have made a call.
3. Concurrency is a second price list. ElevenLabs’ burst rate ($0.16/min above your concurrency limit) doubles the minute cost during spikes. Deepgram’s Growth discount needs an annual commit. Price your peak, not your average.
4. Telephony and tooling are never in the headline. The SIP trunk, the phone numbers, the CRM lookups the agent performs mid-call, the engineering to wire it all — none of the four includes these. A managed service like the ones in our AI phone answering services comparison rolls them into one bill at a higher per-minute rate; that premium is the price of not building.
Understanding how reasoning-token pricing works helps here: on the unbundled platforms, a chatty agent that “thinks out loud” through a large reasoning model can quietly cost more in tokens than in voice minutes.
Who should pick which
| If you… | Lean toward | Why |
|---|---|---|
| Want the lowest per-minute AI fee and can tolerate a week-old product | Mistral Voxtral 3 | $0.04/min all-in is unmatched; EU hosting and self-deployment for regulated data |
| Need one vendor, one invoice, no assembly | Deepgram Voice Agent API | $0.075/min bundled; the longest production track record of the four |
| Need the most natural-sounding voice for a customer-facing line | ElevenLabs | Voice quality is its moat; compare our ElevenLabs vs PlayHT vs Murf breakdown for the TTS side |
| Already build on OpenAI and have engineering in-house | OpenAI Realtime API | GPT-Live 1 from $0.05/min; you own the stack and the savings |
Who should not build a voice agent at all
Most small businesses should not build. If you want an AI answering phones next week — not a platform to assemble — buy the managed service: it charges more per minute because the assembly, the edge cases and the 2 a.m. failures are its problem. Our comparison of AI phone answering services covers that buy, including how the billing unit (per call vs per minute vs per unique caller) decides the monthly bill.
Two compliance notes before wiring an agent to a phone line. Recording calls triggers consent law, which differs by jurisdiction — see our guide to recording consent laws. And a voice agent handling customer data belongs in your AI usage policy: who it may talk to, what it may promise, where audio is retained. Mistral’s retention controls and self-hosting help here; the other vendors’ defaults deserve a read of the fine print.
What we cannot confirm yet
The honesty list: no independent benchmark of Voxtral 3 or Eleven Realtime on production telephone audio exists — all accuracy and latency claims are the vendors’. Mistral never defines its per-minute billing unit, so compare it against Deepgram’s WebSocket-time billing with care. ElevenLabs does not disclose tokens-per-minute for Realtime, so its true cost is an estimate. All rates were read from vendor pages on 23 October 2026; prices change without notice. xAI’s Grok Voice ($0.08/min, model included) is a credible fifth option we did not fully score.
Frequently Asked Questions
What is the cheapest AI voice agent per minute right now?
Mistral’s Voxtral 3 at $0.04 per minute with Mistral Small 4 reasoning — the only headline rate here that includes the LLM. OpenAI’s GPT-Live 1 follows at $0.05/min (backend model and tool usage extra); Deepgram ($0.075/min) and ElevenLabs ($0.08/min) are the next rungs.
Does the per-minute price include the phone call itself?
No. None of the four includes telephony in the headline rate. Budget roughly $0.01 per minute for SIP trunking on top — and remember some platforms bill connection time (hold music and silence included), not just talking time.
Is Mistral Voxtral 3 really half the price of ElevenLabs?
On the AI fee, roughly yes — $0.04 vs about $0.08 — because Mistral bundles the reasoning model and ElevenLabs bills LLM tokens separately. The gap narrows with telephony added, and flips if you compare Mistral’s larger-model tier ($0.08) against ElevenLabs’ Mini. It is also a week-old product with no independent benchmarks yet.
Can I run a voice agent on my own servers?
With Mistral, yes by arrangement — Voxtral 3 supports self-deployment or private cloud, with EU-hosted endpoints as default. Deepgram offers VPC/on-premises at enterprise tiers. ElevenLabs and OpenAI are managed-API only here.
Do I need a developer to set one of these up?
Realistically yes: the four compared here are APIs and SDKs, not plug-and-play products. Without engineering, look at managed phone-answering services or no-code wrappers (Vapi, Retell) that charge a platform fee on top of the model cost.
Are AI voice agents legal to use for business calls?
The technology is legal; recording and disclosure rules depend on where you and the caller are. Many jurisdictions require one-party or all-party consent, and several now require disclosing that the caller is speaking to an AI. Check your jurisdiction’s rules and your AI usage policy before it touches a live line.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
