Skip to content
Voice AgentBible

Voice Agent Weekly, week 43 of 2026: speech-to-speech prices, India's gazette, NYC rules

Two speech-to-speech launches priced per minute, four pricing pages restructured, India's TRAI amendment read in full, and New York City's disclosure rules.

By · 4 min read

Last verified 27 Oct 2026v1.0Published 27 Oct 2026

2026-W43 · 19 Oct 2026 to 25 Oct 2026

Two vendors launched speech-to-speech agent models with published per-minute prices, four pricing pages were restructured between the 2 October snapshot and 27 October, a cluster of preprints measured where full-duplex models fail, we read India's TRAI amendment in the gazette and corrected our India row, and New York City proposed an on-call disclosure rule with a one-step path to a human.

Changelog

WhoWhat changedBefore → afterSeenVerified how
ElevenLabsmodelEleven Realtime v1 Mini and v1 Max speech-to-speech agent models launched; WebSocket only at launch, with an API compatible with the OpenAI Realtime event formatNo speech-to-speech model; Speech Engine for agents at USD 0.08 per minute (snapshot of 29 September) → USD 0.08 per connected minute plus tokens: Mini USD 0.20 input and USD 6.00 output per million tokens; Max USD 1.80 input and USD 9.00 output per million tokens. Vendor-reported time to first audio 0.57 to 0.62 seconds2026-10-27We saw the page
Mistral AImodelVoxtral 3 family (batch and real-time transcription, text-to-speech) and a voice-agent SDK built on LiveKit with SIP and WebRTC; EU, US and global endpoints plus self-hosted deploymentNo per-minute agent price → Transcription USD 0.0035 per minute batch and USD 0.008 real time; text-to-speech USD 13 per million characters; voice agent API USD 0.04 per minute with the smaller language model or USD 0.08 with the larger one2026-10-27Press release
IBMmodelGranite Speech 5.0, a 2 billion parameter speech-to-text and speech-translation model in six languages under Apache 2.0, with keyword biasing lists of up to 2,000 entries; vendor reports 5.06 percent average English word error rate on the Open ASR Leaderboard2026-10-27Press release
GradiummodelPhonon on-device text-to-speech: a 31 million parameter English model under Apache 2.0 and a 40 million parameter five-language model under a non-commercial licence; vendor reports 7 milliseconds to first audio on a laptop CPU2026-10-27Press release
BlockmodelDessa, a full-duplex speech-to-speech research preview that Block says was trained only on conversation audio with no text and no pretrained models; no API, weights, benchmarks or latency figures; waitlist only (23 October)2026-10-27Press release
Sierra Researchreporttau-bench 2.0 adds five languages and an outbound intake domain for voice agents. Across 2,700 calls on three agents, Spanish, Hindi and Portuguese trailed English by 2.4 to 6.4 points on task completion, and two of the three agents missed a caller turn, replied slowly or failed to yield in over 80 percent of calls. The benchmark author also sells a voice agent2026-10-27Press release
CartesiafeatureSonic-2, Sonic-turbo and the 2025-10-27 Sonic-3 version reached their announced sunset date; the pricing page lists Sonic-3.6 onlyOlder Sonic versions callable (sunset announced in the July changelog) → Sunset after 20 October 2026 (we did not observe the endpoints being switched off)2026-10-27We saw the page
OpenAIpricingTwelve-hour prompt-caching option added; cache writes billed at 2.5 times the input-token price (20 October). Realtime and audio rates unchanged2026-10-27We saw the page
TelnyxpricingMid-tier commitment plan restructured on the voice AI pricing page; voice engine unchanged at USD 0.05 per minuteCommitted: USD 500 monthly minimum with volume discounts (snapshot of 2 October) → Growth: USD 2,000 monthly minimum with a 15 percent discount on voice, messaging and numbers2026-10-27We saw the page
Retell AIpricingElevenLabs text-to-speech pass-through split from one flat rate into per-model rates; other voices and platform rates unchangedElevenLabs voices at USD 0.040 per minute (snapshot of 2 October) → USD 0.040 (Flash) to USD 0.100 (v3 and v4) per minute depending on model2026-10-27We saw the page
RimepricingPricing restructured from per-character Starter and Enterprise tiers to Free, Flex, Scale and Enterprise per-minute tiers; a speech-to-speech model (Prism) and a speech-to-text model addedMist v3 from USD 0.03 per 1,000 characters (snapshot of 30 September) → Flex: Prism USD 0.080, Mist v3 USD 0.030, Coda v2 USD 0.050, speech-to-text USD 0.002 per minute (free until 19 November); Scale at USD 100 per month for 20 percent off2026-10-27We saw the page
GladiapricingPublic pricing restructured from Starter and Growth hourly tiers in dollars to a self-serve rate card in euros with per-model async pricing and volume tiers; the currency change may be locale rendering rather than a repricingStarter async USD 0.61 and real-time USD 0.75 per hour (snapshot of 7 October) → Self-serve async from EUR 0.16 (Solaria-3) or EUR 0.27 (Solaria-1) per hour, real time EUR 0.68 per hour2026-10-27We saw the page
GenesysfeatureRelease notes of 19 October: ElevenLabs voices selectable in Genesys Cloud generative text-to-speech; real-time voice transcription for the ServiceNow unified experience; Japanese language models self-enabled by telephony admins; deprecation of some BYOC Cloud TLS ciphers2026-10-27We saw the page
VapifeatureNew provider models the week of 19 October, including Cartesia Sonic-3.6, Claude Haiku 5.5 and Gemini 3.8 Flash; served tier for OpenAI fast mode now shown in call costs. Platform fees unchanged at USD 0.05 per minute hosting (snapshot of 2 October matches 27 October)2026-10-27We saw the page
MicrosoftfeatureTeams Phone Agent gained the Copilot Credit Estimator (19 October) and spam detection in the admin centre (21 October)2026-10-27We saw the page
Synthflow AIotherVoice agents listed on the Zoom App Marketplace for Zoom Phone and Zoom Contact Center; first deployment named as Batteries Plus across more than 700 US locations, with the customer quoting five weeks from kickoff to pilot; no volume, containment or cost outcomes disclosed (20 October)2026-10-27Second-hand
Ofcom (United Kingdom)regulationFined number provider Tismi GBP 437,500 for sub-allocating numbers without due diligence or misuse monitoring, exposing about 366,000 numbers to potential misuse between November 2022 and January 2025; opened an investigation into Xtra Phones UK the same day (21 October)2026-10-27Second-hand
New York City Department of Consumer and Worker ProtectionregulationProposed rules on automated customer-service systems (published 20 October, comments to 19 November): disclosure that the consumer is dealing with an automated system, made at the beginning and in the language of the interaction, and a single-step transfer to a human with an estimated wait, or disclosed alternatives; penalties from USD 525 per violation2026-10-27We saw the page
U.S. District Court, Northern District of CaliforniaregulationIn re Google Cloud Contact Center AI Privacy Litigation: California wiretap claims against the AI vendor dismissed with leave to amend (22 October). The court held that the retailer's recording disclosure, which said recordings may go to service providers, meant two plaintiffs had not plausibly alleged lack of consent, and that none had alleged the vendor's intent2026-10-27We saw the page
TRAI (India)regulationWe read the Third Amendment to the TCCCPR (dated 18 September 2026) in the gazette text. It defines an application-to-person call to include autodialled, robocall and pre-recorded or artificial-voice calls, requires senders to pre-declare A2P use and CLI ranges on the DLT platform, treats undeclared A2P calls as unsolicited commercial communication, and allows a termination charge of up to INR 0.05 per minute except from the 140, 1600 and 1601 series. It contains no duty to announce an AI voice to the called partySite row carried a reported draft duty to declare an AI or synthetic voice at the start of a call → Row corrected: pre-declaration is to the telecom provider, not the caller; commencement staged at 30, 60 and 90 days from gazette publication2026-10-27We saw the page

Pricing watch

Two speech-to-speech launches this week both publish a per-minute agent price, which has been rare. One charges USD 0.08 per connected minute plus language-model tokens at two tiers; the other charges USD 0.04 or USD 0.08 per minute depending on the language model, with transcription and text-to-speech priced separately and an EU-hosted endpoint. Put either into the cost calculator against your call profile before comparing it with a cascaded stack; the token line is the one that moves with how much your agent talks.

Four pricing pages were restructured at some point between the 2 October archived copies and 27 October: one platform raised its mid-tier commitment from USD 500 to USD 2,000 a month with a 15 percent discount, one split a flat text-to-speech pass-through into per-model rates that run from USD 0.040 to USD 0.100 per minute, one text-to-speech vendor moved from per-character to per-minute tiers and added speech-to-text and speech-to-speech models, and one transcription vendor replaced its dollar tiers with a euro rate card. None of these appeared in a release note. If you buy on a plan rather than a rate, re-read the plan page monthly.

Regulatory watch

  • India. We have now read the Third Amendment in the gazette text and rewritten the India row from it. The operative change for voice agents is the application-to-person definition, which covers autodialled, robocall and artificial-voice calls, and the pre-declaration duty: a sender must declare A2P use and the CLI ranges to its telecom provider on the DLT platform in advance, and an undeclared A2P call is treated as unsolicited commercial communication. There is also a termination charge of up to INR 0.05 per minute on A2P calls not made from the 140, 1600 or 1601 series, and AI-based flagging of suspected spam CLIs with barring for repeat flags. What the text does not contain is a duty to tell the called party that the voice is synthetic. Our row said a draft version of that duty had been reported; that was wrong, and the correction is logged.
  • United States, New York City. The proposed rules define an automated customer-service system as one where all or substantially all of the communication is made by technology, and would make it a deceptive practice to use one without disclosing it at the beginning of the interaction in the language being used. A merchant that offers no single-step transfer to a human (say or press "agent", press 0) with an estimated wait would face a rebuttable presumption of a deceptive practice. Comments close on 19 November.
  • United States, courts. The Northern District of California dismissed, with leave to amend, wiretap claims against a cloud vendor whose AI listened to a retailer's customer calls. The retailer's notice said recordings may be shared with service providers, and the court treated that as consent at the pleading stage. It is not a final ruling and it is one district, but it is the first court reading of how far a "this call may be recorded" notice stretches to cover an AI vendor on the line.
  • United Kingdom. Ofcom's fine on a number provider is about sub-allocation hygiene, not AI, but it lands on the layer that decides which numbers your outbound agent presents. Ask your carrier what due diligence it performs on sub-allocated ranges.
  • European Union. No new instrument. The Commission's Article 50 FAQ is worth re-reading on one point: the grace period to 2 December 2026 applies only to the marking of outputs from generative systems placed on the market before 2 August 2026. The duty to tell people they are interacting with an AI system has applied since 2 August and has no grace period.

Research worth reading

Four preprints this week measure where full-duplex speech models break, and together they are a better buyer's checklist than any benchmark score.

  • Memory across a call. A benchmark of 1,094 sessions found that five full-duplex models retained almost none of the facts a caller stated mid-conversation; the best final-state fact coverage was 2.01 percent. For booking and order-taking, that is the whole job.
  • Long calls. On conversations up to an hour, five models eventually degraded into repetition, silence or gibberish, or ran out of context, because they emit a token on every frame even while listening. Vendor demos are four minutes long for a reason. Two training-free fixes are proposed.
  • Deepfake detection over real phone lines. Detectors that work on studio audio degraded badly once the same speech had passed through a real mobile call; the best adapted error rate in the paper was still 22 percent. If a vendor claims caller anti-spoofing, ask what audio it was tested on.
  • Poisoned transcription models. A fine-tuning-time backdoor lets a chosen speaker's natural voice trigger attacker-written transcript text, with success rates over 90 percent on two open model families from as little as two minutes of the target's audio, and the injected text then drove harmful actions in downstream agents. Anyone running a third-party fine-tuned speech-to-text checkpoint under a tool-using agent now has a concrete threat model.

The multilingual benchmark update from a vendor's research group is also worth a look for its method: identical tasks in six languages, with scores for yielding and over-responding, not just task completion. The runs are the vendor's own, and the scores are not comparable with the previous version.

Next week

The regional pages for the United States and India get the items above folded into their regulatory quick cards, and the demo-guide scoring sheet gains a read-back trap based on this week's test.

Site changes this week

  • /compliance/India row corrected from the gazette text (version 1.1). United States, United Kingdom and European Union rows updated with this fortnight's items (version 1.1).
  • /corrections/Correction logged for the India AI-disclosure note.
One thing to test this week. Early in the call, give the agent three facts: a name spelled letter by letter, a six-digit reference and a date. Talk about something else for two minutes, change one of the three ("actually, make that the 14th"), then ask it to read all three back. A benchmark published this week found that current full-duplex speech models kept almost none of the facts stated mid-conversation; a cascaded agent with a transcript usually does better, but check yours.