Voice AI: August's cuts came out of the free tiers

Our first pass over a re-captured pricing page is an automated comparison that pulls out every money figure and sets it against the last capture. It is a good tool. It is fast, it does not get bored on page 19 of 27, and it does not talk itself into seeing a change that is not there.

This fortnight it found nothing worth reporting, and it was wrong twice.

Both changes that actually matter in this window landed on the free tier, and a free tier is usually measured in minutes or characters rather than dollars. So the figure that moved was 3,000 becoming 800, which a money-token comparison reads straight past. You only catch it by reading the page.

Here is everything we can show you a dated capture for, covering 30 July to 16 August 2026.

Speechmatics swapped a perpetual free tier for a one-off credit

This is the biggest change in the window, and it is a change in kind rather than degree.

Until this capture, a Speechmatics account came with a recurring monthly allowance: 3,000 minutes of speech-to-text (50 hours, split 1,200 real-time and 1,800 batch) plus roughly 1M characters of text-to-speech, 2 concurrent real-time sessions, no roll-over. It reset every month. If your usage sat under it, you never paid.

New accounts now get “$100 in credit, no card required”. Once that is gone, you add a card.

The page also states a new unit plainly, which we will give them credit for: “1 credit equals $1, so your credit balance maps directly to the numbers on your invoice.” That is a rare thing on a pricing page. Most vendors invent a token or a unit precisely so you cannot do that arithmetic in your head.

Who this actually hurts. Not the enterprise buyer, who was never on the free tier. It is the developer running a low-volume side project, an internal tool, or a proof of concept that happened to fit inside 3,000 minutes a month. That person was running free forever. Now they run out. Speechmatics has quietly moved from “free for small users” to “free trial”, and those are different products.

We have changed the stored free tier from a free-tier to a credit on our Speechmatics profile, because calling it a free tier would misdescribe it.

Rime cut its free allowance by about 73%

Same shape of change, same blind spot in our tooling.

Rime’s Starter card used to offer 3,000 minutes free to start. It now reads “Get ~800 minutes free (about 800k characters)”. That is roughly a 73% cut, and again, no dollar figure moved, so no price alert fired.

The paid rate is untouched. The card still reads $0.03 per 1,000 characters, exactly as it did on 30 July.

So the trend across both of this fortnight’s real changes is the same: the headline rate is what gets advertised, and the free tier is what gets quietly trimmed. If you are evaluating a platform on its published per-minute cost alone, this is the fortnight that says you are watching the wrong number.

Rime’s July contradiction resolved, and both numbers were real

Last edition we flagged that Rime’s pricing page disagreed with itself. The Starter card said $0.03 per 1,000 characters. The FAQ on that same page said $0.05. We recorded the card as printed, because it is the primary price display and it is what our dated screenshot showed, and we flagged the conflict rather than splitting the difference.

That call has now paid off. The FAQ was updated in this window and reads:

Starter begins at $0.03 per 1,000 characters (about a minute of audio): $0.03 for Mist and $0.05 for Coda.

So neither figure was stale. The $0.03 is the Mist rate, the $0.05 is the Coda rate, and per-model pricing is published again after the mid-2026 redesign stripped it out.

Worth sitting with for a second, because it is the whole argument for how we handle conflicting sources. Averaging the two would have produced $0.04 per 1,000 characters, a rate Rime charges for nothing at all. It would have looked more decisive and it would have been fiction. When a vendor’s page contradicts itself, the honest move is to record what you can see, say which display you trusted and why, and wait.

Deepgram added a model and a deadline, not a price rise

Deepgram’s page moved more than anyone else’s, and none of it is a repricing.

A new text-to-speech model, Flux TTS, is free until 12 September 2026 and $0.0450 per 1,000 characters after that. The Voice Agent API is now on an explicit dated promotion:

Read that table the right way round. The right-hand column is not a price rise; it is the list price we already had stored. The left-hand column is a temporary discount with a published end date. Deepgram is being unusually straight about it by printing both, and you should still budget on the right-hand column, because that is what you pay from 13 September.

Our own first read of this page got it wrong, and it is worth admitting how. Skimming the table, it looked like Aura-2 had risen from $0.030 to $0.045 per 1,000 characters, which would have been a 50% increase on a model plenty of people build on. Reading the rows properly, the $0.045 belongs to Flux, the new model. Aura-2 is unchanged at $0.030. One misread row and we would have published a price rise that never happened.

Three things that looked like news and were not

The point of re-checking 27 pages is as much about the changes you talk yourself out of as the ones you find.

CloudTalk. The hero copy appeared to double the entry bundle, moving from “$99/100 min” to “$99/month with 200 minutes included”. It did not. Our notes have stored $99 for 200 minutes since 23 July, along with a recorded observation that the hero’s “100 min” matched no selectable bundle in CloudTalk’s own dropdown (which offers 50, 200, 500 and 2,500). It was a vendor copy error. What changed is that CloudTalk fixed it. Our data was right and the page caught up.

We nearly published this as a price improvement. What stopped it was checking our own stored notes before writing, which is the only reason to keep notes that detailed.

OpenAI. A large diff, no change. Every rate we store was checked row by row and is identical, including gpt-realtime-2.1 at $32.00 input and $64.00 output per 1M tokens. The churn belongs to the separate Fast mode tier table.

Vapi. The “$50M Series B” banner came off the page header. Prices identical. Funding banners are not pricing news, and it is striking how much of a pricing-page diff is just a vendor’s marketing mood.

Also verified and unchanged: Retell (its language model menu swapped Claude 4.6 Sonnet for Claude 5 Sonnet at the same $0.08/min), Telnyx (call recording, phone numbers and SMS all rose, but those are monthly and per-message line items rather than the call minute), LiveKit (upstream model churn, not LiveKit’s own rate) and My AI Front Desk (a copy rewrite whose new figures are competitor prices it quotes about rivals, so not ours to import).

We also re-scored all 27 platforms

Separately from the pricing sweep, we re-rated every platform in the directory on 16 August.

The reason is not flattering to us. The four sub-scores behind every ranking on this site were set on 30 May and genuinely calibrated only once, on 2 June. Prices moved underneath them across four capture rounds since, and nothing in our automated checks noticed, because those checks tested that a score was a number between 1 and 10, not that it still made sense.

Four scores moved on the evidence. Rime’s value-for-money went up, because a rate that halved and a free tier that shrank still nets out favourably against what it now costs per minute of speech. Trillet, Hermes and ServiceAgent each moved down a point on one axis.

We also fixed three award badges that contradicted the scores printed beside them on the same card, and added automated checks so this particular drift cannot go unnoticed again. The full weightings and method are on our methodology page, which now carries an accurate review date; the old one predated the calibration it claimed to describe.

What to do about it

If you run anything on a Speechmatics free tier, work out today what your monthly usage actually costs at the pay-as-you-go rate, because the $100 will run out and there is no monthly reset behind it.

If you build on Rime’s free allowance, the same applies with more urgency: 800 minutes goes fast, and the cut was not announced anywhere we can find.

If you use Deepgram’s Voice Agent API, budget on the post-12-September figures, not the promotional ones. Standard goes to $0.075 a minute.

If you compare platforms on headline rate alone, this fortnight is the argument against it. Nobody moved a published rate. Two platforms materially cut what you get for nothing.

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