There is no reliable single “price per minute” for a startup voice agent: the total may combine platform hosting, speech recognition, a language model, speech generation, telephony, and optional services. The best-value choice depends on what a call must accomplish, how reliably it must do it, and what the complete workload costs—not on a headline rate or a polished voice demo.
What does a voice AI agent cost per minute?
Start by defining the cost boundary. A hosted-agent price may bundle some services while charging others separately; an API or model price may cover only one component. Before comparing rates, list every charge that applies to the same call scenario.
As an Amazon Associate I earn from qualifying purchases.
- Platform or hosting: agent runtime, orchestration, or access to a hosted platform.
- Speech recognition (STT): converting caller audio to text, if billed separately.
- Reasoning: language-model tokens, tool use, or other backend processing.
- Speech generation (TTS): producing spoken responses.
- Telephony: phone numbers, inbound or outbound minutes, and carrier charges.
- Optional and operating costs: add-ons, concurrency, retention, support, compliance needs, retries, and human escalation.
These line items may be billed by the minute, second, token, character, or another unit. A rate that covers only one part of the stack is not an all-in call price.
Published rates have different service boundaries
| Option | Published rate or offer | What it covers—and what to check |
|---|---|---|
| Vapi usage-based hosting | $0.05 per minute for hosting, on the pricing page accessed October 7, 2026. | Provider model costs are passed through separately, and telephony may also be charged separately. Check the selected STT, reasoning, TTS, carrier, and optional feature rates. |
| Retell AI pay-as-you-go voice agents | $0.07–$0.31 per minute, on the pricing page accessed October 7, 2026. | The range depends on configuration. Use the vendor estimator for the intended model, voice, telephony, and add-ons, then validate against metered calls. The page also lists $10 in free credits and 20 concurrent calls included; verify current eligibility and terms. |
| OpenAI GPT-Live | $0.05 per minute, billed per second, on the pricing page accessed October 7, 2026. | This is a session charge, not a complete agent bill. Backend model and tool use are separate; session duration includes silence. |
| Google Gemini API | Model-specific speech prices; no single comparable rate is stated here. | Identify the model, input and output modality, and free or paid tier. Google’s pricing page lists planned changes effective January 1, 2027 for certain models, so confirm the applicable rate for the budget period. |
| ElevenLabs API | Model-specific speech rates; no single comparable rate is stated here. | Confirm the chosen model and current pricing. The pricing page advertises a startup grant of 12 months free and 33 million characters; treat this as conditional, and confirm eligibility and application terms rather than assuming the offer applies. |
These figures are not directly comparable all-in prices: they refer to different service boundaries and billing models. Rates and offers can change; the figures above reflect pages accessed October 7, 2026.
#1 Best Overall
- Stay present in every scenario: Every conversation is covered, in person, on calls, and online. 4 MEMS + 1 VPU microphones with AI beamforming capture every voice across the room. Smart Dual-Mode Recording switches automatically between phone calls and in-person. The free Plaud Desktop captures online meetings without a bot
- Walk out of every meeting with notes ready to act on: Plaud Intelligence transcribes in 112 languages with speaker labels and turns each recording into action items, decisions, and follow-ups, structured and ready to use. Choose from 10,000+ customizable templates tailored to your role and industry
- AI summary ready before you reach your desk: Auto Transfer moves each recording to the Plaud app automatically, and AutoFlow transcribes and summarizes so your notes are ready before you are back at your desk. Upgrade anytime to Pro (1,200 min/mo) or Unlimited
- Access your AI workspace anywhere: One connected workspace across Plaud Desktop, Plaud Web, and the Plaud mobile app, so your conversations and finished work follow you everywhere
- Your conversations stay private and yours: Compliant with ISO 27001, ISO 27701, SOC 2, HIPAA, GDPR, and EN 18031, with zero data used to train AI models. Trusted by 2.5M+ professionals, including legal, medical, and business professionals handling sensitive information
How to read a worked estimate
Vapi’s documentation gives an illustrative four-minute GPT-Live call estimate of about $0.48 plus telephony, using rates stated as of September 30, 2026. Its example assigns $0.20 to voice time, $0.20 to the platform, and about $0.08 to the reasoner. It assumes particular token use and eight delegations; it is an illustration, not a quote. Actual reasoning cost can change with prompt size, tool results, and how often the agent delegates work. Verify current rates and measure your own calls.
Which voice AI platform gives a startup the best value?
There is no evidence here for a universal quality winner, and price alone cannot establish value. Choose the comparison that matches your operating model, then score every option against the same workload and success criteria.
Choose a hosted platform when integration speed is the priority
Compare hosted options such as Vapi and Retell on setup effort, orchestration controls, test and analytics tools, telephony arrangements, concurrency, and the full metered cost. Retell’s pricing page lists templates, analytics, transcripts, simulation testing, webhooks, and API access, but whether those features fit a particular startup depends on its needs and current plan terms. A hosted platform can reduce integration work; it does not remove the need to check which components and operating requirements are billed separately.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsRank #2
- Stay present in every scenario: Every conversation is covered, in person, on calls, and online. 4 MEMS + 1 VPU microphones with AI beamforming capture every voice across the room. Smart Dual-Mode Recording switches automatically between phone calls and in-person. The free Plaud Desktop captures online meetings without a bot
- Walk out of every meeting with notes ready to act on: Plaud Intelligence transcribes in 112 languages with speaker labels and turns each recording into action items, decisions, and follow-ups, structured and ready to use. Choose from 10,000+ customizable templates tailored to your role and industry
- AI summary ready before you reach your desk: Auto Transfer moves each recording to the Plaud app automatically, and AutoFlow transcribes and summarizes so your notes are ready before you are back at your desk. Upgrade anytime to Pro (1,200 min/mo) or Unlimited
- Access your AI workspace anywhere: One connected workspace across Plaud Desktop, Plaud Web, and the Plaud mobile app, so your conversations and finished work follow you everywhere
- Your conversations stay private and yours: Compliant with ISO 27001, ISO 27701, SOC 2, HIPAA, GDPR, and EN 18031, with zero data used to train AI models. Trusted by 2.5M+ professionals, including legal, medical, and business professionals handling sensitive information
Choose direct APIs when stack control matters
With direct integrations, evaluate speech recognition, reasoning, synthesis, telephony, observability, and engineering effort as separate decisions. This can provide more control over component choice, but comparing it fairly means adding each component’s usage charges and accounting for the work of integrating, monitoring, and maintaining the stack.
Choose on workload fit, not on a “best voice” label
Define whether your product needs clear speech in noise, natural-sounding short responses, reliable interruptions, domain vocabulary, multiple accents, or sensitivity to emotional delivery. A voice demo can help assess voice fit, but it cannot establish task completion, recognition accuracy, turn-taking, or cost per successful call.
How should a startup test voice quality and latency?
Run a fixed evaluation that reflects the product’s actual users and call environment. Give every candidate the same script, call route, network region, turn policy, and definition of success; otherwise, a difference in results may come from the test setup rather than the platform.
Rank #3
- YOUR AI PERSONAL ASSISTANT FOR EVERYDAY PRODUCTIVITY: More than a voice recorder, Pocket works as your AI personal assistant to capture, transcribe, and summarize meetings, calls, and ideas instantly. Core features are included out of the box, with optional advanced tools available for power users.
- ONE-TAP RECORDING FOR REAL-LIFE MOMENTS: Capture meetings, phone calls, and in-person conversations instantly with a simple tap, no typing, no interruptions, just effortless note-taking anywhere you go.
- SMART AI INSIGHTS & ORGANIZATION: Pocket automatically turns recordings into clear summaries, key action items and structured conversation maps so you can quickly review what matters without digging through audio.
- TURN CONVERSATIONS INTO ACTION WITH “ASK POCKET”: Don’t just record, understand. Instantly ask questions across your meetings, extract key insights and generate next steps in seconds. All grounded in your recordings, so answers stay accurate and reliable.
- MAGSAFE COMPATIBLE FOR SEAMLESS USE: Easily attach Pocket to your iPhone or other MagSafe compatible devices for convenient, hands-free recording on the go. Perfect for capturing meetings, calls, and ideas without needing to hold your device.
Build representative test calls
- Include the target languages, accents, names, and domain vocabulary.
- Test noisy and low-bandwidth audio, short and long utterances, and realistic speech interruptions.
- Exercise the real business task, including tool calls and the recovery path when something goes wrong.
- Use the same network region and telephony route across candidates where practical.
Measure user-perceived delay, not just model speed
Record end-to-end time-to-first-audio and turn latency from the application, including networking and application overhead. Capture latency distribution and slow-tail behavior rather than reporting only an average. Model inference time is not the same as the delay a caller experiences. ElevenLabs’ latency documentation makes this distinction explicitly: “When diagnosing latency in your application, measure from your application, not from API benchmark figures.”
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →ElevenLabs describes its Flash models as faster, with less quality headroom than its larger, more expressive Eleven v3 family. That is vendor guidance about its own models, not an independent ranking across providers. Test the specific voices and model versions you intend to deploy.
Score the complete interaction
- Task completion and critical errors: Did the agent reach the intended outcome, and did it make a consequential mistake?
- Recognition: How often did speech recognition mishear the target accents, names, or terms in the target environment?
- Turn-taking: Can users interrupt naturally, and does the agent recover without losing context?
- Tool calls: Do integrations succeed when the agent needs to act, not just speak?
- Human ratings: Ask representative listeners to rate naturalness, intelligibility, and fit for the intended product or brand.
- Economics: Calculate cost per call and per successfully completed task, including retries and human escalation.
What quality risks matter for sensitive voice use cases?
If a decision depends on how something is said as well as the words spoken, test that capability directly. A June 2026 preprint, Real-Time Voice AI Hears but Does Not Listen, evaluated four realtime voice systems across three consequential scenario types and reported that systems often acted on words while discounting vocal delivery. The authors described improvements from prompting as partial and inconsistent. This finding is a caution for those scenarios, not proof that every voice system or use case has the same limitation.
Rank #4
- Plaud Intelligence: Capture conversations in 112 languages and generate accurate transcripts with the Plaud App and Web. Plaud Intelligence uses leading models like GPT-5.5, Claude Sonnet 4.6, and Gemini 3.1 Pro to transform raw audio into structured insights. Choose from over 10,000 professional templates to generate mind maps and to-do lists, turning hours of discussion into immediate clarity
- Multiple Ways To Wear With Included Accessories: Adapt Plaud NotePin S to any workflow instantly with four included accessories. Wear your device effortlessly as a necklace, wristband, clip, or pin. Plaud NotePin S features a dedicated physical record button for precise, tactile control. Stay professional and keep your intelligence within reach all day
- Enterprise-grade Privacy: Built to the highest standards with ISO 27001/27701, SOC 2, HIPAA, GDPR, and EN18031 compliance. Every conversation is secure and protected. It is the trusted choice for creative, medical, and business professionals handling sensitive info
- Multimodal Input & Multidimensional Summaries: Capture audio, type notes, add images, and press/tap to highlight for richer context with multimodal input. Press the record button to mark key moments in real time. Plaud transforms a single conversation into multiple perspectives, providing faster, clearer insights, and unifies these inputs to deliver role-specific summaries that reflect your intent and priorities
- Lightweight Power and Peace of Mind: Weighing only 0.61 oz, Plaud NotePin S delivers 20 hours of continuous recording and 40 days of standby time. Store up to 64GB of audio locally, ensuring you capture every insight even without an internet connection
For emotion-sensitive or otherwise consequential workflows, include delivery-specific test cases and retain appropriate human review. Do not treat a transcript, a generic voice-quality score, or prompt instructions as proof that a system interpreted vocal cues correctly.
How can a startup compare offers without overestimating savings?
Build a monthly estimate around expected call volume, call length, and the actual workflow. Apply each vendor’s billing unit to the same scenario, then include all required components and costs generated by unsuccessful calls.
- Set the workload: estimate monthly calls, average and long-call duration, expected concurrency, and the share needing retries or escalation.
- Map the bill: record the rate and billing unit for platform, STT, reasoning, TTS, telephony, and necessary add-ons.
- Check limits and terms: confirm concurrency, minimum packages, retention, support, compliance needs, and whether any credits or grants apply to your company.
- Run the evaluation: measure task success, errors, latency, and call cost under representative conditions.
- Compare cost per success: divide full workload cost by successfully completed tasks, and include retries and human handoffs in the total.
- Recheck rates before committing: confirm pricing, offers, and effective dates for the models and tiers you plan to use.
For context, Creative Genius reports a production comparison of 12,400 calls across eight client production numbers and 11 platforms over 90 days, spanning inbound sales, appointments, service intake, and support. It is a commercial integrator’s report, not a controlled neutral benchmark, so its findings should not be treated as a universal leaderboard. Use it as context for the kinds of deployments represented, not as a substitute for testing your own workload.
What should the final comparison include?
Keep one scorecard for each candidate and record both what the vendor publishes and what your own evaluation establishes. That prevents a published component rate or a polished sample from standing in for evidence about the complete product.
Quick Recap
- Billing unit, included services, pass-through charges, and all-in cost for the test scenario.
- Latency measurement start and end points, plus typical and slow-tail results from the application.
- Language and accent coverage, speech naturalness, and intelligibility for target users.
- Interruption handling, turn-taking, recognition errors, and tool-call success.
- Task completion, critical error rate, and cost per completed task.
- Integrations, telephony, concurrency, reliability, data retention, and compliance requirements.
- Support terms, credits, grants, and other startup access offers, with eligibility and expiry confirmed.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
