This comparison is a pricing snapshot checked on October 17, 2026. Voice-agent price pages change, and each company's calculator depends on configuration. The goal is not to declare one service universally cheapest. It is to show what a quote contains, identify the parts that are easy to miss, and give you a repeatable way to estimate the same workload across OpenAI Realtime, Vapi, and ElevenLabs.
What counts as a voice-agent minute?
A customer may think a minute is simply the time between answering and hanging up. A supplier can measure that interval differently from a model. A speech-to-speech model may bill input and output audio tokens only when a response is generated; a hosted agent platform may bill every minute an assistant runs; a voice provider may include a fixed number of call minutes in a subscription and charge for additional minutes. Transcription, a separate large language model, text-to-speech characters, telephony transport, recording retention, and concurrency can add further cost.
Before comparing vendors, write down six assumptions: inbound or outbound, browser or telephone, average session duration, fraction of time the person speaks, fraction of time the agent speaks, and expected concurrent sessions. Add the exact model, voice, transcriber, tool-call rate, and whether conversations use a long prompt or retained history. Without those assumptions, “cost per minute” is a marketing shorthand rather than a comparable bill.
OpenAI Realtime: token-priced responses
OpenAI's Realtime API does not bill a conversational session as a single flat number of minutes. It accrues usage when responses are created and charges based on input and output tokens across audio, text, and image modalities. A response may be triggered manually or by voice activity detection. The conversation is reused as input on subsequent turns, so later replies can include a larger history; stable matching input can qualify for prompt caching. Input transcription, when enabled, is a separate usage path.
The October 17 rate card lists gpt-realtime-2.1 audio input at $32 per million tokens, cached audio input at $0.40 per million, and audio output at $64 per million. Its text rates are $4 input, $0.40 cached input, and $24 output per million tokens. The gpt-realtime-2.1-mini rates are $10, $0.30, and $20 for audio input, cached audio input, and audio output; its text rates are $0.60, $0.06, and $2.40. Check the live OpenAI API pricing table before putting numbers into a quote because model names and prices can change.
OpenAI's cost guide estimates one user audio token per 100 milliseconds and one assistant audio token per 50 milliseconds, with small variations from special tokens. As a deliberately simplified upper-usage illustration, one minute of continuous user audio is about 600 input audio tokens, while one minute of continuous assistant audio is about 1,200 output audio tokens. At the listed base audio rates, that works out to about $0.0192 of user-audio input plus $0.0768 of assistant-audio output for a full two-sided minute on gpt-realtime-2.1, or about $0.096 before text tokens, transcription, and any other separately billed service. The same rough token arithmetic on mini is about $0.006 plus $0.024, or $0.03. This is not a per-minute quote: real speech has pauses, turns, cached history, text, optional transcription, and tokenization variation.
OpenAI fits teams that want the provider's live audio model and can own the session server, tools, transport, and usage accounting. The advantage of one token-based rate card is flexibility to reason from actual response usage. The tradeoff is that you must collect and reconcile those usage details yourself. A timer started at call answer cannot produce the exact Realtime model cost.
Vapi: platform hosting plus selected components
Vapi's current pricing page shows a platform hosting fee of $0.05 per minute in its 1,000-minute calculator example. The calculator then breaks a sample stack into transcriber, intelligence model, voice, and transport. In the displayed configuration, Deepgram transcription is estimated at $0.0095 to $0.0099 per minute, an OpenAI model at $0.0077 to $0.0452 per minute, and an ElevenLabs voice at $0.0146 to $0.0238 per minute. Added together with the $0.05 hosting fee, those components total approximately $0.0818 to $0.1289 per minute before transport. These are calculator estimates for the shown selections, not a guaranteed rate for every assistant.
At 1,000 minutes on those same shown ranges, the arithmetic is $50 hosting, $9.50 to $9.90 transcription, $7.70 to $45.20 model, and $14.60 to $23.80 voice: about $81.80 to $128.90 before any chargeable transport or optional success package. For US telephony examples on the pricing page, Twilio transport is listed separately at $0.008 per inbound minute and $0.014 per outbound minute. Vapi WebRTC, WebSockets, and Vapi Telephony / SIP are shown as free transport in the page's current table; external carrier pricing can differ. The live Vapi calculator and package details should be rechecked with the providers you actually select.
Vapi's documentation says the Usage Only option charges a per-minute hosting fee plus model, voice, transcriber, and telephony costs. Optional Core, Pro, and Premier Success Packages add support, service commitments, capacity, or governance without replacing usage charges. The current page lists Core at $29 per month. Pro is described as 10% of Vapi hosting spend with a $999 monthly minimum, and Premier is custom. Legacy customers may retain earlier terms. The included call-line count, raw-data retention, and phone-number allowance differ by package, which matters if your deployment needs concurrency or longer support logs.
Vapi is useful when you want a platform that composes providers and gives you one place to configure assistants, calls, and usage estimates. Do not compare only the $0.05 platform component with a model's full token price; they are different rows of the eventual bill. Review the cost shown for completed calls, the selected provider rates, and which party bills the phone carrier.
ElevenLabs: monthly minute plans, plus model choices
ElevenLabs prices ElevenAgents through monthly plans with included call minutes and concurrency. The public page currently lists Free at $0 with 15 minutes and 4 concurrent calls; Starter at $6 per month with 75 minutes and 6 concurrent calls; Creator at $22 with 275 minutes and 10 concurrent calls; Pro at $99 with 1,238 minutes and 20 concurrent calls; Scale at $299 with 3,738 minutes and 30 concurrent calls; and Business at $990 with 12,375 minutes and 40 concurrent calls. Introductory discounts can appear on the same page, but they expire and should not be used as a long-term cost assumption.
For minutes beyond the included amount, the page lists $0.08 per call minute. It lists burst pricing at $0.16 per minute when calls exceed the plan's concurrency limit, and $0.003 per text message. The key qualifier is that external LLM usage is billed separately based on the model; the pricing page says applicable model fees are deducted from ElevenLabs credits. If you connect an external telephony provider, that provider bills its carrier charges directly. ElevenLabs says it does not add its own telephony fee, while the external carrier can still charge for the connection.
This plan structure is easy to understand if your usage stays within a predictable monthly envelope and the included concurrency suits your traffic. A company planning 1,000 minutes should compare the $99 Pro plan's 1,238 included minutes with the model configuration and any external carrier charges, rather than multiply all 1,000 minutes by $0.08 and ignore the plan bundle. The ElevenAgents pricing page is the source for current plan minutes and overage rates.
Side-by-side: the bill's unit and the hidden variable
| Provider | Primary pricing shape | Variable that changes the estimate | Good fit when |
|---|---|---|---|
| OpenAI Realtime | Per-response input and output tokens, by modality and model | Speech ratio, turn count, retained history, cache hits, optional transcription | You want a direct Realtime integration and can operate session accounting |
| Vapi | Per-minute hosting plus selected model, voice, transcriber, and transport | Provider combination, prompt and tool size, call length, phone route, success package | You want a hosted composition layer and want to choose providers per component |
| ElevenLabs ElevenAgents | Monthly plan with included minutes and concurrency; overage by call minute | Included-plan fit, external LLM usage, concurrency bursts, optional carrier | You value a bundled agent experience and your traffic fits a published minute tier |
This table is about billing shape, not product quality. Audio quality, interruption handling, voice licensing, supported languages, observability, data location, retention, uptime terms, and tool control can matter more than a small difference in a calculator. Confirm each of those in the product documentation and contract that applies to your account. A public list price may not be the rate in a negotiated enterprise agreement.
Build a like-for-like 1,000-minute estimate
- Fix the workload. Use one defined channel, average call length, inbound/outbound mix, speech ratio, and expected concurrency for all three options.
- Specify each voice stack. For OpenAI, name the Realtime model and whether transcription is separate. For Vapi, select a transcriber, model, voice, and carrier. For ElevenLabs, choose a plan and whether the model or telephony is external.
- Include plan costs and overage. Multiply the Vapi components by the expected call minutes, and add its hosting or support package. For ElevenLabs, compare the monthly plan allowance, unused included minutes, overage, and any model or carrier costs.
- Calculate OpenAI from usage scenarios. Estimate audio and text input/output tokens by response, then run low, typical, and high conversation-history cases. Apply cached rates only to the part likely to be cached; do not assume every repeated prompt is a cache hit.
- Add the operational cost. Include application servers, logging, storage, phone numbers, retries, compliance, support packages, and engineering time if they are relevant to the purchase decision.
Publish the assumptions next to the estimate. A range such as “$80 to $130 per 1,000 minutes on this Vapi configuration, before transport and optional support” tells the reader more than a false universal cents-per-minute value. Keep the exact date and provider selections with the calculation so an updated price page does not rewrite last month's planning exercise.
Track cost and customer usage as separate measurements
A provider's cost model answers what the service charged your company. Your product meter answers how much of a defined quantity a customer consumed. You might record call count, Realtime audio tokens, or hosted minutes. Those quantities can support reporting and later pricing decisions, but a quantity is not a provider invoice and it does not automatically calculate margin.
For Realtime, preserve the provider's per-response usage and the rate card used for the estimate. For Vapi, save its completed-call cost breakdown when available and retain the component configuration. For ElevenLabs, distinguish the monthly plan charge, included minutes, additional call minutes, external model usage, text messages, and carrier bill. Use a stable session identifier in your own database to reconcile those records. If you need customer-level aggregation, decide which resource identifies the customer before ingestion; do not assume a usage endpoint accepts arbitrary labels it does not document.
UsageBox can aggregate a deliberately chosen numeric quantity through meters. It does not apply OpenAI, Vapi, or ElevenLabs rates, combine carrier invoices, calculate gross margin, or invoice your customer. Keep your rating and payment workflow in the billing system that owns those jobs. If your product has a product-specific billable unit, meter that unit separately from your internal provider cost so you can change pricing without erasing what the customer actually used.
Which one should you shortlist?
Shortlist OpenAI Realtime when a direct integration and response-level usage control matter, and you have engineering capacity to manage server credentials, WebRTC lifecycle, tool permissions, and accounting. Shortlist Vapi when provider composition and a configurable hosted layer are worth a visible per-minute platform charge. Shortlist ElevenLabs when its monthly minute bundles, concurrency tiers, and voice-agent workflow match your traffic shape, and you have included any external model or carrier costs.
Then run the same evaluation script against each design. Compare cost for a representative sample of real call lengths, successful task completion, tool errors, latency, transfer rates, and behavior under interruption. Do not build a decision around only the provider's best-case sample call. A lower-cost voice that causes users to repeat themselves may create more total minutes; a slightly higher unit rate can be better if it reduces failed turns. Measure these outcomes in your own pilot rather than assuming a vendor's calculator has accounted for them.
These prices are a dated snapshot, not contractual advice. Check the linked official rate pages again before launch or procurement, and record the chosen rate card with every financial estimate. The comparison above aims to make the bill legible: model usage, platform hosting, call minutes, provider add-ons, transport, and customer pricing are different costs and should remain identifiable.
For current plan details, calculation methods, and provider charges, check the linked official pricing pages again before using this snapshot for a purchase or customer quote.
Related guide
For an implementation that collects Realtime usage and keeps the provider's dollar rate separate, see the OpenAI Realtime build and metering walkthrough.