Voice agents, Nov 2026 to May 2027: November 17 can retire two Gemini Live previews. January 20 retires gpt-realtime.

The published minute is seven different bills

A voice-agent minute is a label vendors put on different stacks. One includes the language model. One is the session clock, silence included. One is hosting. One is the overage after a free allotment. Ranking them as a single price requires the column that says what was left out. The chart is that ranking with the caveat in the caption. The table is the number. ^12911121820

ElevenLabs sells concurrency, at a flat minute

Additional minutes are $0.08 on Free, Starter, Creator, Pro, Scale, and Business. Included minutes run 15, 75, 275, 1,238, 3,738, and 12,375. Concurrent calls run 4, 6, 10, 20, 30, and 40. Monthly fees on the page are $0, $6 after a $1 first month, $22 after an $11 first month, $99, $299, and $990. Text messages are $0.003. The FAQ says burst pricing lets agents handle up to three times the plan's concurrency at $0.16 a minute. The page does not say what happens past that ceiling. A worked bill for 1,000 and 10,000 three-minute calls, including the plan-fee arithmetic, is the cost note. This page does not repeat it. ¹²²

OpenAI's two meters, in one minute of audio each way

The cost guide says user-message audio is 1 token per 100 ms and assistant audio is 1 token per 50 ms. A minute of caller audio is 600 tokens. A minute of agent audio is 1,200 tokens. The guide also says the entire conversation is sent on each response, so a later turn costs more than the first, and that voice activity detection filters empty input audio. The figures below are this page's multiplication for one minute of each, before that history and before the special tokens the guide says will move the count by a small amount. They are not an OpenAI invoice. ²⁴

GPT-Live-1 stays on the session clock. $0.05 a minute, billed per second, with no round-up to a whole minute, and the backend model extra. A silent hold is billable on GPT-Live-1. On gpt-realtime-2.1, with voice activity detection on, empty input audio is not an input token unless the client adds it. Retell's speech-to-speech table, a third party's price for OpenAI's models, lists GPT Realtime 2.1 at $0.38 a minute and GPT-Live 1 plus the backend at $0.05. That $0.38 is Retell's bundle. It is not the token arithmetic above. ²⁴¹⁹

LiveKit's HTML calculator, on the phone-call toggle for Build/Ship, displayed $0.0479 a minute. The lines were session $0.01, telephony $0.01, an LLM at $0.0014, speech-to-text at $0.0075, speech at $0.0090, and observability at $0.01. They sum to the displayed total. The total row did not name the models. On the same page's per-minute lists, Gemma 4 31B is the only LLM at $0.0014. Two AssemblyAI streaming models are at $0.0075. AssemblyAI TTS 1 is at $0.0090. That is a match against the list. It is not a label the calculator printed. ¹³

Speech-to-speech and the cascade are both on sale

The architecture choice is which process owns the turn. A cascaded stack runs speech-to-text, then a language model, then text-to-speech, with a separate turn-taking model deciding when the caller has finished and when a barge-in should cut the agent off. A speech-to-speech model takes audio in and speaks audio out. Tool calls still have to land somewhere. OpenAI and ElevenLabs each sell both shapes. ⁵¹⁷

The key design choice is how speech connects to reasoning and tools: a continuous conversation with a separate backend, a single voice model, or a pipeline you control stage by stage.

OpenAI's guide names the three. GPT-Live is full duplex: the live model listens and speaks at the same time and delegates reasoning and tools to a backend, either your own agent or an OpenAI Responses model. The Realtime API puts speech, reasoning, and tool use in one session, and the session handles interruptions. The chained pipeline is speech-to-text, your agent, then text-to-speech, which is the path that lets you store the transcript and run a policy check before anyone speaks. The guide's example chained agent uses the model id gpt-6-astra. That id is a text model in the sample, not a voice rate. ⁵

ElevenLabs describes the same split from the other direction. ElevenAgents coordinates a speech-to-text model, a language model you choose, a text-to-speech model, and a proprietary turn-taking model. The overview says agents can also use a Realtime speech-to-speech model. Speech Engine is the orchestration layer for a team that already has a text agent. The Realtime models are also a standalone API, and the page points at an OpenAI Realtime wire-protocol compatibility guide. The overview says the text-to-speech model is low latency. It does not print a millisecond. ¹⁷

Nobody published one end-to-end millisecond

End-to-end latency is the wait from the end of the caller's turn to useful audio, including speech recognition, the model, tools, speech synthesis, and the network. The pages read for this brief do not publish that number on a shared method. OpenAI tells you to measure audible response latency at the median and the 95th percentile, with the caller, the recording, the backend model, the prompt, the transport, and the grader held fixed, and to time an acknowledgment such as 'I'm checking' separately from the answer. LiveKit's pricing card says the media transport delivers voice and video worldwide in under 250 ms. That sentence sits under media transport. It is not time-to-first-audio for the agent. Deepgram's acquisition post says Daily's network is engineered for ultra-low latency and cites 75 points of presence. Both phrases are Deepgram's. This page did not measure a call. ⁵¹⁰¹²

Turn-taking is a product surface with its own price on at least one rate card. LiveKit's markdown lists audio end-of-turn detection and adaptive interruption as free for agents hosted on LiveKit Cloud. Self-hosted agents get a development allowance and then pay, on Ship and Scale, $0.50 per 1,000 end-of-turn requests and $1.20 per 1,000 adaptive-interruption requests. The turn-detection doc says a realtime model with server-side turn detection decides the interruption itself, and that Gemini Live keeps turn-taking on its own signals, so this SDK's client-side interruption options do not apply. That sentence is LiveKit's account of Gemini. It is not a line this page found on Google's pricing or deprecations pages. ¹²¹⁶

Evals are a line item, when a vendor prices them

OpenAI's guide says to score task completion, tool arguments, interruptions, unwanted silence, and incomplete sessions separately, and to move from synthetic single turns to recorded human speech to a simulated multi-turn caller. It points at a cookbook. It does not print a price for the eval. LiveKit does. On Ship and Scale, text simulations are $0.02 a turn and audio simulations are $0.05 a turn. Build includes 350 text turns and 60 audio turns. Cartesia's pricing page says evaluations are free of charge for a limited time only, and it does not name the end date. A limited-time free eval is a watch item, not a permanent zero. ⁵¹¹¹²

Pipecat leads stars. The description still says Daily.

Deepgram announced on October 29, 2026 that it is acquiring Daily, the company that makes Pipecat. The post says the framework stays open source and free under its existing BSD V2 license, that an independent Pipecat Foundation is being established, and that the Daily team joins Deepgram. It also introduces a Real-Time Context Standard, co-developed with Pipecat, and says a Voice Agent Platform is forthcoming. The post tells readers to get in touch for platform updates. It does not name a general-availability day. This page did not fetch the standard's specification file. ¹⁰

Keeping Pipecat open was fundamental to our decision to join Deepgram.

The gap is 2,020 stars, Pipecat ahead. LiveKit has 1,032 more forks. Pipecat's latest release, v1.15.0, was published at 05:50 UTC on October 29, the same calendar day as the acquisition post. LiveKit's latest release, 1.8.11, was published November 5, six days before this check. The repository description on Pipecat still reads "Maintained by Daily and the community." Thirteen days after the announcement, the API field had not been updated. The license field is SPDX BSD-2-Clause, which matches the announcement's "BSD V2" in name. This page did not diff the LICENSE file. ¹⁰¹⁴¹⁵

Stars and downloads disagreed the day before. This site's November 10 census, which this page did not re-sum, counted 4,583,195 PyPI downloads without mirrors for livekit-agents and 1,468,981 for pipecat-ai from October 10 through November 8. That morning's stars were 15,161 and 17,168. The install leader was LiveKit. The star leader was Pipecat, and it still is, by a slightly wider gap. A download is a file request. A star is a bookmark. Deepgram's post calls Pipecat the world's most popular open-source voice agent framework. That is Deepgram's sentence. The counts above are the ones this site pulled. The census is LiveKit Agents had 4,583,195 downloads. ¹⁵²¹

The phone line is its own meter

A voice agent that never leaves the browser is a WebRTC session. A voice agent that rings a handset needs a phone number and a SIP trunk, or the vendor's own telephony. Those charges are small next to a talkative model and large next to a careless comparison, because some stickers include them and some tell you the carrier will bill you directly. ^112181920

LiveKit's included SIP minutes are 1,000 on Build, 5,000 on Ship, and 50,000 on Scale. US local inbound minutes are a thinner allowance: 50, 100, and 1,000, then $0.01. A toll-free inbound minute on Ship and Scale is $0.02, and the number is $2 a month. Concurrent agent sessions are 5 on Build, 20 on Ship, and up to 600 on Scale, starting at 50. Those caps are the production limit a pilot does not see. ElevenLabs buys the same thing with the plan: 4 concurrent calls on Free, 40 on Business, and a burst rate rather than a hard stop the page is willing to describe. Bland Start stops at 100 calls a day and 10 concurrent. The page says the cap is the cap. ¹¹²²⁰

Retell's add-ons sit on top of the stack and are easy to omit from a pilot quote. A knowledge base is an extra $0.005 a minute. Batch calling is $0.005 a dial. Branded calling is $0.10 an outbound call. Advanced denoising, safety guardrails, and PII removal are $0.005 a minute each. The first 20 concurrent calls are included, then $8 a concurrent call a month. None of that is in the $0.11 named stack. ¹⁹

What is dated from November 2026 through May 2027

Announced means a linked vendor page names the day. Likely means a document aims the work at that month and does not lock the day. Watch means a later check should confirm the item or drop it. A month with no dated voice-agent event stays empty in the announced column. The general six-month outlook covers product filings and regulation. This table is voice agents only. ³⁷⁹

Announced

November 17, 2026 is the earliest date Google prints for two Live API preview models, gemini-3.1-flash-live-preview (released March 11, 2026) and gemini-2.5-flash-native-audio-preview-12-2025 (released December 12, 2025). The recommended replacement on both rows is gemini-3.8-live, released September 15, 2026, with no shutdown date on the stable Live API table. The same earliest date is on gemini-3.1-flash-tts-preview, a text-to-speech preview, replaced by gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts. The page's note applies to the whole set of tables: the shutdown column is the earliest a model might be retired, and Google says it will communicate the exact day. Six days after this check is the earliest, not a confirmed off switch. ⁷

The price of that move, on the audio rates Google prints, does not change for the 3.1 Flash Live preview. The pricing page puts gemini-3.8-live, gemini-3.8-live-extended-thinking, and gemini-3.1-flash-live-preview on one paid table: $3.00 per million audio input tokens or $0.005 a minute, and $12.00 per million audio output tokens or $0.018 a minute. The older native-audio preview is a separate block at the same $3.00 and $12.00 token rates, without the per-minute equivalents printed beside them. Migrating the preview id is a model-id change. It is not, on these tables, a new audio rate. ⁶⁷

January is firmer because OpenAI states a removal day, not an earliest day. On January 6, 2027, tts-1, tts-1-hd, and gpt-4o-mini-tts leave the API. The recommended replacement in that section is gpt-realtime-2.1-mini, which is a realtime model, not a drop-in text-to-speech endpoint. On January 20, 2027, gpt-realtime, the August 28, 2025 snapshot, gpt-4o-realtime, gpt-realtime-mini, the December 15, 2025 mini snapshot, and gpt-4o-mini-realtime leave. Replacements are gpt-realtime-2.1 and gpt-realtime-2.1-mini. The same January 20 table also removes gpt-audio, gpt-4o-audio, and the mini audio ids. Developers were notified July 20, 2026. The pricing page still lists the models being removed, at the older audio rates of $32 and $64, or $10 and $20 for the minis. The rate card has not dropped the ids ahead of the shutdown. ²³

December 31, 2026 is the other printed day, and it is a credit, not a voice-agent price. Deepgram gives $1 of Flux text-to-speech credits per $1 spent, up to $500, through that date. The Voice Agent rates beside it have no end date. ElevenAgents Campaign Manager, published November 9, is already in the past relative to this check. The docs page this site read on November 10 did not add a separate fee. The product note is What is ElevenAgents Campaign Manager?. ⁹

Likely

Likely, in the sense this site uses on the other outlooks, is a document that aims work at a month inside the window and does not lock the day. No page in this set does that for a voice agent. Deepgram's October 29 post says the Voice Agent Platform is forthcoming and points readers at a contact form. Forthcoming is not March, and it is not May. Cartesia's free evaluations have no end date, so they are not a likely expiration in a named month. This column stays empty on purpose. ¹⁰¹¹

Watch for

Watch Google's exact-date notice for the two Live previews. If November 17 passes and the ids still answer, the table did what it said it might do: it named the earliest day, not the day. Watch January 6 and January 20 on OpenAI, which are removal days, not earliest days. Watch whether Deepgram publishes a platform date, whether the Pipecat GitHub description drops the word Daily, and whether any minute in the rate table grows an effective date. Enterprise contact-center contracts from vendors this brief did not open, including Genesys, NICE, Five9, Amazon Connect, and Google Contact Center AI, are not in the watch list because no public rate card from them was read. A private price is not a blank this page can fill. ³⁷¹⁰¹⁵

What this means for builders and buyers

The work between November 11, 2026 and May 2027 is a migration of model ids and a cleanup of quotes. It is not a wait for a cheaper minute. The pages that publish a minute did not publish a day when it changes. The pages that publish a day are shutting model ids down. ¹³⁷

Builders on gemini-3.1-flash-live-preview or gemini-2.5-flash-native-audio-preview-12-2025 should move the id to gemini-3.8-live before they need the exact-date email. The audio token rates on the shared 3.8 table match the preview's token rates. The per-minute equivalents are printed on the 3.8 block. Builders on tts-1 or gpt-4o-mini-tts should read the January 6 replacement as a product change: the named replacement is gpt-realtime-2.1-mini. Builders on gpt-realtime, gpt-4o-realtime, or gpt-realtime-mini should move to gpt-realtime-2.1 or the mini before January 20. The audio rates of the 2.1 pair, $32 and $64 or $10 and $20 per million, match the rates still printed for the ids that are leaving. The bill changes shape if the call pattern changes. The list price of a token, on this card, does not. ²³⁶⁷

Builders choosing a stack should write down who owns turn-taking and who owns the tool call. A cascade lets you read the transcript and reject a tool call before speech goes out. Full duplex lets the caller talk while the backend works, and the backend is a second meter on GPT-Live-1. A realtime session hides the boundary and bills tokens that grow with the history. LiveKit will host the turn detector for free and charge for it if you self-host past the allowance. ElevenLabs will sell you the cascade, with its own turn-taking model, or a Realtime model inside the same agent product, at the same $0.08 voice minute, with the LLM still extra. ¹⁵¹²¹⁷

Buyers should ask four questions and refuse a quote that answers with one number. Which model id, and what does the vendor's deprecations page say about it? Which of the seven minutes in the chart is the quoted minute, and where are the model, the voice, and the phone? What is the concurrency cap, and is the overage a higher minute, a rejected call, or a $10 line? Who pays for the eval, per turn or per minute, after the pilot? The 1,000-call worksheet is the invoice arithmetic for ElevenLabs, Vapi, Retell, and Bland. The ElevenLabs directory listing is the product page. The general agent calendar is the six-month outlook. ¹⁵²²

Framework buyers have a narrower job. Pin livekit-agents 1.8.11 or pipecat 1.15.0, or a newer tag you have read. Keep Pipecat if the BSD terms on the repository still match the October 29 promise when you next review the license. Do not treat 17,194 stars as evidence the framework places the call, and do not treat 4,583,195 downloads as evidence it is cheap to run. Both numbers are real. Neither one is a minute rate. ¹⁴¹⁵²¹

Methodology

This page did not field a survey, recruit a panel, interview vendors, place a call, or charge a card. It copies public pages and APIs and labels every arithmetic step. There is no hidden sample of practitioners and no measured latency of our own.

What was counted

Prices, plan fees, concurrency caps, and telephony rates are strings from the vendor pages listed in the sources, fetched November 11, 2026. The HTTP Date or response Date is in the source line. Vapi's response Date is November 10, from a cache HIT. Retell's Last-Modified is November 7. GitHub figures are stargazers_count, forks_count, license.spdx_id, pushed_at, and the latest release's published_at. 17,194 minus 15,174 is 2,020. OpenAI audio minutes use 60 divided by 0.1 seconds, which is 600 caller tokens, and 60 divided by 0.05 seconds, which is 1,200 agent tokens, times the per-million rate. 600 times $32 per million is $0.0192. 1,200 times $64 per million is $0.0768. The sum is $0.096. The mini pair sums to $0.030. Vapi's low end sums to $0.0818. The high end sums to $0.1289. Retell's named stack sums to $0.11. LiveKit's six displayed lines sum to $0.0479. ^241314151819

What a word means

Announced means the linked page names a day. Likely means the linked page aims the work at that month without a day. No page in this set did that, so the likely column says so. Watch means the next check should confirm or drop the item. Earliest possible, on the Gemini table, is Google's phrase, and this page keeps it. A vendor sentence that says low latency or ultra-low latency is quoted as a vendor sentence. It is not converted into a millisecond.

What this does not measure

It does not measure end-to-end latency, task success, barge-in accuracy, or answer quality. It did not re-sum the PyPI downloads in the November 10 census. It did not open Genesys, NICE, Five9, Amazon Connect, Google Contact Center AI, Sierra, or Decagon rate cards, so enterprise contact-center contracts are outside the table. It did not open Bland's billing doc, so the Scale plan that doc carries is not a row here. It did not switch Retell's country dropdown. US Twilio is the first option, value 0.015, and the span next to the select displayed 0.015. It did not fetch the Real-Time Context Standard specification, ElevenLabs' standalone Realtime price if it differs from the Agents page, or a Gemini shutdown email. The deprecations fetch followed a redirect to AI Studio. The shutdown table was in that response. ^7192021

Forward-looking cells stay labeled. A month stays empty when no source names it. Deepgram's claim that Pipecat is the most popular open-source voice framework is their claim. The star counts are the API's. If you cite a number from this page, cite the source number and the clock time.