Skip to content
Voice AgentBible

Voice Agent Weekly, week 42 of 2026: promo prices end, FCC consent order, call report

Introductory TTS prices end, a speech-to-speech model ships, FCC consent-revocation rules take effect 25 November, UK nuisance-call complaints hit a five-year high.

By · 3 min read

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

2026-W42 · 12 Oct 2026 to 18 Oct 2026

The first changelog issue, covering the fortnight since launch. An introductory text-to-speech price ended on 12 October, a speech-to-speech model shipped at an unchanged per-minute rate, the FCC's consent-revocation order takes effect on 25 November, Pennsylvania's narrower calling window began on 18 October, and the ICO reported automated-call complaints at a five-year high.

Changelog

WhoWhat changedBefore → afterSeenVerified how
ElevenLabspricingIntroductory 72 percent discount on Eleven v4 and v4 Turbo text-to-speech ended on the API price listv4 at USD 0.022 and v4 Turbo at USD 0.011 per 1,000 characters, labelled 72 percent off until 12 October (page snapshot of 29 September) → v4 at USD 0.08 and v4 Turbo at USD 0.04 per 1,000 characters2026-10-27We saw the page
ElevenLabsfeatureAudio effects option added to Text to Speech and Text to Dialogue at no extra cost; speech-to-text gained a turn-prediction commit strategy2026-10-27We saw the page
xAImodelGrok Voice 3.0 speech-to-speech model released through the Speech to Speech API; agents on the grok-voice-latest alias were migrated automaticallyGrok Voice 2.0 at USD 0.08 per minute → Grok Voice 3.0 at USD 0.08 per minute (vendor states latency under one second; no independent measurement)2026-10-27Press release
Fish AudiomodelDrama 3 open-weight text-to-speech model: 35 billion parameters (about 3 billion active), 83 languages, community licence free for commercial use under USD 500,000 annual revenue, no cloning without consentNo open-weight Drama 3 → Weights on Hugging Face; hosted API listed at USD 50 per million characters, USD 10 promotional rate until 12 November2026-10-27Press release
SpeechmaticsfeatureMelia 1 speech-to-text available in real time with automatic language detection across 56 languages and word-level code switching (13 October); Oak 1 medical model for batch (16 October)n/a → Price list shows Melia 1 at USD 0.24 per hour batch and USD 0.40 real time; Oak 1 at USD 0.30 batch and USD 0.75 real time2026-10-27We saw the page
OpenAIfeatureGPT-Live sessions can act while the user is still speaking (15 October); prompt-to-voice custom voices opened to all API customers, creation free, with nine new voices in German, French, Hindi, Japanese and Korean (16 October)Audio and realtime rates on the price list (snapshot of 1 October) → Same rates on 27 October; gpt-live-1 at USD 0.05 per minute, gpt-live-transcribe at USD 0.017 per minute2026-10-27We saw the page
TelnyxpricingSix preset avatars for Voice AI Assistants entered beta with a per-minute charge announced for 1 November 2026 (13 October); Programmable Voice available in Mumbai with local SIP connectivity (16 October)Avatars not offered → Free in beta, then USD 0.08 per minute from 1 November 20262026-10-27We saw the page
TwiliofeatureVoice Insights tags each call as HD Audio (AMR-WB or Opus) or Downsampled (9 October); Voice Integrity registers business details with the three major US call-analytics providers (beta, 16 October). The US voice price page still reads current as of August 20262026-10-27We saw the page
MicrosoftfeatureTeams Phone Agent, generally available since 30 September, gained 42 voices (14 October). Billing is usage based through Copilot Credits; without credits the line falls back to a standard auto attendant. Not supported for use in India2026-10-27We saw the page
Microsoft AImodelFirst streaming transcription model (MAI-Transcribe-2-Streaming, 60 languages) and two text-to-speech models released on Microsoft Foundry (1 October, before the launch issue)n/a → Streaming transcription at an introductory USD 0.54 per hour; Voice-2.1-Flash at USD 15 and Voice-2.1 at USD 22 per million characters2026-10-27Press release
Automation AnywhereotherDefinitive agreement to acquire Boost.ai, a conversational and voice AI vendor; terms undisclosed; close expected in the fourth quarter of 2026 subject to approvals (7 October)2026-10-27Press release
Transaction Network ServicesotherAcquires netnumber, a number-intelligence, portability, routing and identity-verification provider; terms undisclosed (15 October)2026-10-27Second-hand
FCC (United States)regulationReport and Order on consent revocation (FCC 26-67), adopted 30 September and released 1 October; published in the Federal Register on 26 October with an effective date of 25 November 2026A revocation request applies to all robocalls from the caller; delayed effective date of earlier rules extended to 31 January 2027 → Callers may treat a revocation of informational robocalls as applying only to that category and may designate one clearly disclosed exclusive revocation method; a revocation sent in response to telemarketing still revokes all future marketing robocalls2026-10-27We saw the page
PennsylvaniaregulationAct 47 of 2026 amends the Telemarketer Registration Act; in force from 18 October 2026. Telephone solicitation now includes voicemail, ringless voicemail and text; robocalls to any line need prior express written consent; synthetic or computer-generated messaging used to deceive is an express violationSolicitation window 8 a.m. to 9 p.m. → No solicitation on Sundays, or after 7 p.m. or before 9 a.m.2026-10-27We saw the page
Colorado Department of LawregulationInterim draft rules for automated decision-making and conversational AI services (6 October; hearing 26 October; proposed effect 1 January 2027) would require an audible AI disclosure at the start of the first interaction each day, at least every three hours in a continuous interaction, whenever the user asks, and as an intermittent audio disclaimer on products without a screen2026-10-27We saw the page
ICO (United Kingdom)regulationComplaints about automated calls rose from 1,374 in July to 2,178 in August and 3,684 in September 2026, the highest since 2021; the ICO links the rise to automated and AI-enabled marketing calls and says it is investigating (9 October)1,374 complaints in July 2026 → 3,684 complaints in September 2026; 15 penalty notices totalling GBP 2,370,000 issued since April 20252026-10-27We saw the page
NextPhonereportAI receptionist call report from one vendor's inbound traffic, December 2024 to October 2026: 14.1 percent of business-line calls arrived after hours and converted at 23.6 percent against 18.6 percent in hours; 1.6 percent of callers asked whether they were talking to a robot and converted at 21.9 percent against 19.2 percent; calls from self-declared AI callers rose from 0.7 to 4.0 per 10,000 (12 October)2026-10-27We saw the page
ElevenLabsoutageStatus page lists SIP call disruptions for agents in the EU (6 October) and globally (8 October), and US endpoint timeouts and text-to-speech failures after a traffic spike (13 October); non-SIP agent conversations were reported unaffected; durations not stated2026-10-27We saw the page
TwiliooutageVerify one-time codes and Programmable SMS delivery failed in several regions for about six and a half hours (7 to 8 October); voice products were not listed as affected2026-10-27We saw the page

What this issue covers

The launch issue promised that twenty pricing pages would go under watch. This issue is the first result, and because the launch issue carried no vendor changes, it also includes the dated changes from the first week after launch and the days just before it. Every row in the changelog carries the page we read and the date we read it. Where a before value comes from an archived copy of the page, the row says so.

Pricing watch

Of the twenty pages, eight were unchanged between their late-September or early-October archived copy and 27 October: Bland AI, Cartesia, Google (Gemini API), Inworld, LiveKit, OpenAI, SignalWire and Twilio. One page (Azure Speech) renders its prices in the browser and could not be compared.

The changes that did land fall into three kinds. Introductory prices ending: the v4 and v4 Turbo text-to-speech models returned to list price on 12 October, which is the kind of change that shows up on next month's invoice rather than in a release note. New charges announced ahead of time: avatars on one platform are free in beta and become a per-minute charge on 1 November. New models at unchanged rates: a speech-to-speech model moved to version 3.0 at the same per-minute price and migrated existing agents automatically, which is convenient and also means your agent's behaviour changed without a deploy on your side. If you run acceptance tests, re-run them.

Regulatory watch

  • United States, federal. The FCC's consent-revocation order was released on 1 October and takes effect on 25 November 2026. The order lets a caller treat a revocation of informational robocalls as applying only to that category, and lets a caller designate a single, clearly disclosed way to revoke consent. A revocation sent in response to a marketing call still revokes consent to all future marketing robocalls, and the interactive opt-out mechanism for artificial-voice calls that carry an advertisement is unchanged. The order contains no AI-specific provisions; an AI voice is an artificial voice, as it has been since February 2024.
  • United States, Pennsylvania. Act 47 of 2026 took effect on 18 October. The solicitation window is now 9 a.m. to 7 p.m. with no calls on Sundays, robocalls to any line need prior express written consent, and synthetic or computer-generated messaging used to deceive is an express violation. The act does not require an on-call AI disclosure, whatever some summaries say.
  • United States, Colorado. The Attorney General's draft rules of 6 October set out when an audible "this is AI" disclaimer would have to be played: at the start of the first interaction each day, at least every three hours in a continuous interaction, whenever the user asks, and as an intermittent audio disclaimer on a product without a screen, in the language of the conversation. It is a draft with a hearing on 26 October and a proposed effective date of 1 January 2027.
  • United Kingdom. The ICO's figures are the first time a regulator has tied a measured complaint spike to AI-enabled calling. The numbers are complaints, not call volumes, and the ICO does not split AI calls from recorded messages. For anyone running outbound campaigns into the United Kingdom, the direction is clear enough.

Reports worth reading

One vendor published a call dataset that is worth reading with its caveats attached. The after-hours figures (14.1 percent of calls, converting better than in-hours calls) match what buyers in home services and dental tell us, and the disclosure figure is the useful one: 1.6 percent of callers asked whether they were talking to a robot, and those callers converted slightly better, not worse. It is one vendor's traffic, mostly US law firms and local services, with no total call count and a transcription-model change in August, so treat the percentages as a starting point for your own count.

Two preprints this fortnight attack the same problem from different angles: native full-duplex models spend compute on every audio frame, including while listening. One uses rolling masked diffusion to predict several frames per pass and reports 1.5 to 1.8 times speedups within an 80 millisecond interaction interval. The other separates "when to speak" from "what to say" so the reply can be reasoned about in parallel. Neither has shipped code yet; both are relevant to the cost per concurrent call of speech-to-speech agents.

Incidents

Two platform incidents are worth noting for how they fail rather than that they failed. One vendor's SIP-connected agents were disrupted on 6 and 8 October while its web and WebRTC agents kept working, so a phone-first deployment needs its own monitor on the telephony leg. A separate SMS and one-time-code delivery failure on 7 and 8 October did not touch voice, but any agent that verifies a caller by text code was blind for six hours. Check which of your call flows depend on a channel you are not monitoring.

Next week

The regulatory table gets its first revisions: the India row is rewritten from the gazette text, and the United States, United Kingdom and European Union rows pick up this issue's items.

Site changes this week

  • /newsletter/Subscriptions are now handled on the site's own server with double opt-in; confirmations and unsubscribes work from the email.
  • /about/About and editorial policy moved to version 1.1; the site is published by Voice Agent Bible Research.
One thing to test this week. Halfway through the call, ask "am I talking to a robot?". A good agent answers in one plain sentence that it is an automated assistant, then carries on with the task. A bad one dodges, denies, or restarts its greeting. Write down the exact words it used; Colorado's draft rules and New York City's proposed rules would both make the answer to that question a legal requirement.