Voice AI in July 2026: every verified price and product change

Between our 15 June and 11 July 2026 pricing captures: Synthflow removed all public pricing and now starts at $30,000 a year, Frontdesk killed its free tier, Speechify replaced pay-as-you-go with subscriptions, Cartesia raised two plan prices, and OpenAI renamed its realtime models. Three vendors joined the directory. Every change below is screenshot-verified.

That list is the whole format of this piece, the first in a monthly series. We capture every pricing page in the directory on a dated sweep and compare it with the last one; June’s ran on the 15th, July’s on the 11th, so every old and new figure below sits between two screenshots we hold. Outside news carries its announcer and its date.

The big one: Synthflow pulled its prices

Sometime between 15 June and 11 July, Synthflow deleted its public price list. Our June capture still showed the full pay-as-you-go table: $0.09/min for the voice engine, per-model AI charges, $0.02/min managed telephony, roughly $0.15 to $0.24/min all-in, and a $2,000 a month white-label add-on. The July capture shows one enterprise card. It reads “Enterprise contracts start at $30,000 annually”, scopes the deal around call volume, and carries a contact-sales button under a row of compliance badges (SOC 2, GDPR, HIPAA, ISO 27001).

The signup path went with the price list: every button on the site now routes to a talk-to-sales form, and the homepage still offering a free build reads as leftover copy. The direction was signposted on 23 June, when Synthflow announced a partnership with AVANT, an enterprise technology distributor.

What it means if you are buying. If you were a self-serve Synthflow customer, the product you priced no longer exists to buy; the closest like-for-like builds are on the Synthflow alternatives page. If you are an enterprise buyer, you now negotiate against a disclosed floor, and the withdrawn public rates are your benchmark: at the old $0.15 to $0.24 a minute all-in band, $30,000 a year bought roughly 125,000 to 200,000 minutes (30,000 ÷ 0.24 and 30,000 ÷ 0.15). Run under about 10,000 minutes a month and the new floor prices you out of the old economics; run well past it and a flat contract can genuinely beat metered rates. Our position: a vendor that removes public pricing has decided small buyers are not worth serving, and small buyers should take the hint.

Every price change, verified

Frontdesk first, because the rename hides the sharper change. The vendor formerly listed as My AI Front Desk now brands as Frontdesk, and its perpetual free tier (20 voice minutes a month) is gone. The new Basic plan is $20 a month ($16 paid annually) and includes zero voice minutes: 40 texts and 10 website-chat conversations, no calls. For a voice product, that is a strange entry plan. The practical floor for phone calls is now Business-in-a-Box at $99 a month with 200 minutes, unchanged from June, which works out at 99 ÷ 200, about $0.50 a minute effective.

Cartesia’s rises are real but narrow. Pro went $4 to $5 and Startup $39 to $49, both up roughly a quarter, and the annual-billing toggle has disappeared from the page. The rates that drive call bills did not move: the agent minute stays $0.06 and telephony $0.014. So this lands on hobbyists and small teams paying the subscription, not on per-call economics.

The ElevenLabs row is a clarification rather than a change, included because the wrong number is circulating. Creator’s regular price is $22 a month for 121k credits; the $11 you may have seen is a first-month 50 per cent promo shown on the plan card. Budget $22 from month two. The agents rate held at $0.08/min all month.

And one number moved in the buyer’s favour: Speechmatics grew its free tier from 2,400 to 3,000 minutes a month, split 1,200 real-time and 1,800 batch, plus roughly a million characters of text-to-speech. That is 25 per cent more free transcription, no strings found.

Product and model changes

Speechify’s developer platform is now SpeechifyAI, living at speechify.ai (the old pricing URL redirects there permanently). The flat $10 per million characters pay-as-you-go and the $5,000 enterprise annual minimum are gone, replaced by four subscription tiers running from Free (50k characters plus 60 agent minutes a month) to Scale at $499, each with its own overage rate. The bigger news for this directory is the new Voice Agents API: an all-in rate from $0.07 a minute, the page promising “no passthrough, no token math”, covering the model, the transcription, the voice and the phone routing in one figure. Overage steps down from $0.075 on Starter to $0.068 on Scale, with Enterprise from $0.06.

Retell rebuilt its model menu. The AI-brain line now runs from GPT 5 nano at $0.003 a minute to GPT 5.5 at $0.16 ($0.32 on the fast tier), with GPT 5.4 at $0.080, GPT 4.1 at $0.045, Claude 4.6 Sonnet at $0.08 and Gemini 3.5 Flash at $0.081. Watch the ordering trap: Gemini 3.5 Flash costs nearly double GPT 4.1. ElevenLabs voices joined the voice menu at $0.040 a minute against $0.015 for the standard options, and the page’s all-in banner now reads $0.07 to $0.31 a minute. If you costed a Retell build before July, your numbers are stale.

Retell also un-gated its compliance paperwork, and this one saves real money. The BAA and DPA (the health-data and data-processing agreements that US healthcare and EU-facing buyers need signed) are now self-serve at no extra fee on any plan, where a BAA previously meant an enterprise conversation. Two honest wrinkles: Retell’s pricing page still lists “Custom BAA” under Enterprise, so the vendor’s own pages disagree, and the same compliance docs state Retell does “not currently operate services within the European Union”. Read that line carefully if EU data residency is the reason you wanted the DPA.

OpenAI renamed its realtime models on 6 July: gpt-realtime-2 became gpt-realtime-2.1, the mini likewise, at identical token rates ($32 per million audio-input tokens, $0.40 cached, $64 per million out; the mini at $10, $0.30 and $20). New in 2.1 is image input at $5.00 per million tokens ($0.50 cached), so a realtime agent can now look at a picture mid-conversation. gpt-realtime-1.5 dropped off the pricing page entirely, so if your build pins model names, check them.

AssemblyAI made the same kind of move: Universal-3 Pro became Universal-3.5 Pro on 23 June at unchanged prices ($0.21 an hour for pre-recorded audio, $0.45 for live streaming). The new realtime model advertises 18 streaming languages with mid-sentence switching, where the retired model’s page advertised 99+. We read that as a more honest count, not a capability collapse, but treat advertised language counts sceptically. Speaker diarization (telling voices apart in a transcript) is now a paid add-on: $0.02 an hour on pre-recorded, $0.12 on streaming.

Speechmatics launched Melia 1 on 17 June, a multilingual transcription model covering 55+ languages that can follow a speaker who switches language mid-sentence. It is batch-only for now (you upload files; no live calls), and the advertised “as low as $0.129/hr” applies only if you agree to share your audio for model training; decline that and the rate is higher. Ten free hours a month to test it. The partner and affiliate programme has also quietly gone; the old partner URL now lands on a startup-credits scheme.

The market moved around the platforms

Four things happened outside the pricing pages.

OpenAI launched GPT-Live on 8 July: two full-duplex voice models (full-duplex means the model listens and talks at the same time instead of waiting for your turn to end). They replace Advanced Voice Mode in ChatGPT, and OpenAI’s system card says complex work gets delegated to GPT-5.5 in the background. What you cannot do yet is build on it: there is no API pricing, just a sign-up list. For developers, gpt-realtime-2.1 remains the product OpenAI actually sells.

Retell shipped Conductor on 29 June, a copilot that builds and edits voice agents and shows every proposed change on the affected workflow step for a human to approve before it reaches production. Retell’s launch numbers say 2.5 times faster to a first production agent and about 60 per cent less build time; those are the vendor’s own figures, not ours. It is live inside Retell now with no separate price published.

xAI opened its Voice Agent Builder beta on 1 July, a no-code builder on Grok Voice. Per xAI’s announcement, it charges $0.05 per minute of audio with voices included and no platform fee, plus $0.01 a minute for telephony on a number xAI provisions free. That would put a phone agent at 0.05 + 0.01, about $0.06 a minute all-in, under the cheapest bundled rates in the directory. The caveats: it is a beta running a single in-house model with no menu, and the figures come from the announcement rather than a hands-on bill.

And the money story: Bloomberg reported on 2 July that ElevenLabs is in early talks over an employee share sale at roughly a $22 billion valuation, about double the $11 billion of its February Series D. Early talks, reported not confirmed; nothing changes on your bill. We note it because a valuation doubling in five months tells you what investors think this market is about to do.

New to the directory

Three listings joined during the sweep. One paragraph each, with the wrinkles we logged.

Hermes is a white-label operations layer for agencies running voice agents on Retell or Vapi underneath: Starter $149 a month with 300 minutes and 3 client workspaces, Business $399 with 1,000 and 7, Agency $699 with 1,650 and 10, all with $0.21 a minute overage. Its own integrations page says $0.24 for the same overage; we recorded the pricing page’s $0.21 and flagged the conflict. It is in private beta (100 operator spots, 60 days free), so the numbers are promotional until public launch.

Trillet AI is also white-label agency tooling: Studio $99 a month with 100 minutes, Agency $299 with 300, Enterprise custom, with a flat $0.12 a minute overage and $0.05 a minute once a call transfers to a human. Two flags from our capture: the homepage shows a SOC 2 Type II badge while the vendor’s own trust centre says the audit is in progress, and it is an Australian company whose pricing page names no currency (its terms point to USD, which we recorded).

Fish Audio is a produced-audio text-to-speech house rather than a call platform. The API charges $15 per million UTF-8 bytes across its current models, which works out near $0.015 per 1,000 English characters, roughly $0.021 a spoken minute by our derivation. Its S2 model weights were open-sourced back in March 2026; that is new to the directory rather than new news.

What to do with July

Four concrete moves, depending on where you sit.

If you are a self-serve Synthflow customer, price a migration now rather than at renewal; the alternatives page is the shortlist. If ElevenLabs Creator is in your budget sheet, the number is $22, not $11. If you run on Retell, re-cost the build against the new model menu in the calculator and, if you touch health data, sign the now-free BAA. And if you were eyeing Frontdesk’s $20 plan for phone cover: zero voice minutes; the calling plan is the $99 one.

Three weeks until the EU disclosure duty

The closer, because it is the one change this month with a fine attached. From 2 August 2026, Article 50 of the EU AI Act requires callers to be told they are talking to an AI system, clearly and at the latest at the first interaction, with fines up to €15 million or 3 per cent of worldwide turnover for skipping it. If your agents take or make calls touching the EU and the greeting does not disclose, you have three weeks. The full breakdown, with the regulation text quoted and dated, is in our EU AI Act guide.

That is July. The August edition of this changelog lands after our next audit sweep, in the same format: every change bracketed by two dated screenshots, and nothing claimed that we cannot show you a capture for.