Voice AI in August 2026: every verified change

Three platforms moved their prices in the second half of July, out of 27 we re-captured. Telnyx rebuilt its pricing page and added commitment tiers at $500 and $5,000 a month. Rime cut its published rate 40%, from $0.05 to $0.03 per 1,000 characters, and left its own FAQ quoting the old number. OpenAI renamed a service tier and put a price on GPT-Live transcription for the first time. Everything else held.

That list is the whole format of this piece. We capture every pricing page in the directory, set the new screenshot against the last one, and only report what the pair can show you. This edition brackets 23 July against 30 July, with the wider July window where a change started earlier.

The change with a deadline attached

Start here, because it is the one with a fine on the end of it and it is not a price.

On 20 July 2026 the European Commission adopted its final guidelines on the Article 50 transparency obligations. Reference C(2026) 5054 final, 51 pages. The duty itself was already known: as of 2 August 2026, people have to be told when they are talking to an AI system. What was missing until now was the Commission’s own reading of the one word the whole thing turns on, “obvious”.

The guidelines settle it, and not in the direction the industry was hoping. The exemption “should be interpreted restrictively”. The provider has to “assess and demonstrate” that the artificial nature of the interaction is obvious, so the burden sits with the platform, not the caller. And on interactions that are “auditory in nature”, the guidelines list the factors that make an interaction less obvious, naming “vocal tone (robot voice vs. genuine human-sounding voice)”.

Read that last one twice if you are shopping for a voice agent. The better your agent sounds, the weaker your legal cover. The naturalness is the product and it is also the thing that removes your excuse.

We downloaded and read the document rather than the coverage of it, and the full breakdown with paragraph references sits in our AI Act guide. If your agents touch EU callers and the greeting does not disclose, you are out of time.

Telnyx rebuilt its pricing page

Same URL, different page. The $0.05 per minute voice engine did not move, but almost everything around it did.

The rate is now sold as “No fine print”: orchestration, hosted speech-to-text and hosted text-to-speech in the one number, with language-model tokens at Telnyx GPU prices and telephony at carrier cost on top. Underneath, three commitment tiers replaced the old two.

Note what those tiers are not. They do not gate features. Telnyx’s own line is “Pick your commitment, not your features”, and every plan runs the same platform, models and network. You are buying a discount and a human, not capability.

The telephony components got itemised too: a $0.002/min Voice API platform fee with SIP billed on top, SIP inbound from $0.0032/min and outbound from $0.005/min, call recording at $0.0012/min, US local numbers at $0.25/mo. Our stored telephony estimate of $0.007 a minute still balances against that, so the all-in maths on the Telnyx profile is unchanged.

One thing we did not adopt. The page’s own estimator outputs $0.056 a minute for a production agent, and the closing banner reads “Ship a voice agent for about six cents a minute.” That is the vendor’s arithmetic with the vendor’s cheapest model. We keep our own derived low of $0.06 rather than publish a supplier’s estimate as our figure, and you can see the difference for what it is: a rounding argument, not a disagreement.

The part that needs a flag is the new compliance section. It reads “Compliance is included. Not an add-on”, and it points at rivals charging “up to $2,000/month extra for HIPAA”. That is marketing copy, and it does not change our HIPAA tick, because Telnyx’s own help centre says the opposite. That article, last updated 29 January 2026 and re-checked during this sweep, says Telnyx relies on the HIPAA conduit exception and does not sign business associate agreements. A marketing section claiming regulated-industry readiness and a support article declining to sign a BAA are not the same statement. We record the conflict and leave HIPAA unticked.

Rime cut its rate 40%, and its own page disagrees

The clearest price move of the month, and the messiest.

Rime’s pricing card read “STARTING AT $0.05 / 1K characters” on 23 July. On 30 July the same card reads $0.03. That is a 40% cut on the only published number the company gives you, and it puts the single rate back where the old per-model Mist rate sat before the mid-2026 redesign removed the per-model lines.

Then scroll down the same page. The FAQ answer to “How is Rime priced?” still says “Starter begins at $0.05 per 1,000 characters of generated speech.”

So the page contradicts itself, and we are not going to average two numbers to make the problem go away. We record the card, because the card is the primary price display and it is what the screenshot shows, and we flag the FAQ as stale. If you are budgeting on $0.03, get the rate card in writing first.

The knock-on is real. At $0.03 Rime now undercuts Cartesia’s roughly $0.035, which moves it up a row in the cost table on our narration roundup and takes a 10-minute script from about 45 cents to about 27 cents. Only Speechify at $0.01 and Fish Audio at about $0.015 are cheaper.

We did not re-rate Rime’s value-for-money score on the back of this, and that is deliberate. A 40% cut would justify moving it. Re-ranking a board on a number the vendor’s own page disputes would not.

OpenAI renamed a tier and priced GPT-Live

Two changes, and the smaller-looking one matters more.

The rename first: “Priority processing was renamed Fast mode on July 30, 2026”, with both service_tier: priority and fast accepted. The rates on that tier moved a long way down, but that tier is separate from the standard realtime rates, and it is worth being precise about which is which, because a quick diff of the page makes it look like headline cuts. We compared the model table row by row. Every standard rate we track is unchanged: gpt-realtime-2.1 still $32.00 in, $0.40 cached, $64.00 out per 1M audio tokens, the mini still $10.00 / $0.30 / $20.00, translate still $0.034/min, whisper still $0.017/min.

The change that actually matters is two new lines in the table. gpt-live-transcribe now carries $0.017 a minute, and gpt-transcribe $0.0045 a minute. Neither was on the page a week earlier.

GPT-Live launched on 8 July as a ChatGPT feature with API access promised “soon” and no published pricing, which is exactly what we recorded at the time. It now has a price on part of it. That is the first hard number on the full-duplex model everyone has been waiting to build against, and it is cheap enough to be interesting.

The four that looked like news and were not

Half the changes our diff flagged were banners. Worth naming, because a banner swap is what most “vendor announces” coverage is built on.

• Fish Audio replaced its “Limited Time Offer - 50% OFF YEARLY” banner with an anniversary announcement reading “ONE YEAR. $52M SEED. 8M+ BUILDERS.” The discounted annual prices the old banner advertised are unchanged at $66 / $450 / $8,988. So the countdown ended and the discount did not, which is the third label this same offer has carried since we started watching it. On the raise itself, note the vendor banner says $52M and press coverage on 28 July reports $50M; we record both and reconcile neither.
• Speechmatics retitled a pricing FAQ from “How does Model Training work?” to “What is the model-training discount programme?”, now described plainly as “an opt-in programme that takes 33% off Speech to Text rates”. Clearer wording for the same data-for-discount trade we wrote up in the discount piece. Rates unchanged. Its startup-credits banner is absent from this capture.
• Trillet rewrote its pricing copy and dropped its agency economics calculator. Plan prices identical: $99 Studio, $299 Agency, Enterprise custom.
• Frontdesk rotated a customer testimonial. That is the entire change.

What did not change, which is most of it

Eleven of the 27 pricing pages were byte-identical to the previous capture. Nine more differed only in page furniture with no dollar figure moved, three of those by a zero-character delta, meaning the same length with different bytes, which is rotating marketing text.

That ratio is the honest headline of any month. Most platforms do nothing, most of what looks like news is a banner, and the way you tell the difference is having last month’s screenshot.

What to do with this month

Four moves, by who you are.

If you sell into the EU and your agent does not disclose, stop reading and fix the greeting. The duty is live as of 2 August, and the Commission has said in writing that sounding human makes your exemption weaker, not stronger.

If you are on Telnyx and spending over $500 a month, ask about the Committed tier. It gates no features, so if you are already past the minimum the discount is free money you are not claiming.

If you buy narration by the character, re-price Rime, but get the $0.03 confirmed in writing before you move budget, because the vendor’s own FAQ still says $0.05.

If you have been waiting to build on full-duplex, GPT-Live transcription now has a published rate. Not the whole model, but the first real number.

Our price index carries every current rate with its capture date, and the calculator runs your own volume against all of them. The September edition lands after the next sweep, same format: two dated screenshots per claim, and nothing asserted that we cannot show you.