Deepgram Raises Its Agent Rate 34% on 12 September

Deepgram's pricing page prices its Voice Agent API at $0.056 a minute through 9/12, then $0.075.

That is a 34% increase [computed], published in advance, with a named date on it. Fourteen days from now.

It is the most honest form of price change in this category and it will still break a budget built in August. Anyone who costed a deployment from this page in the last month and did not read the four characters after the word through has a number that expires on 12 September 2026.

The same page carries a second dated finding pointing the other way, and this one takes an archive to see.

What follows: the two rates that matter, the strikethrough that turns out to be genuine, what the September increase costs at three volumes, the reseller selling Deepgram's own model below Deepgram's own list, why this vendor carries no computed all-in figure here, and what recognition is actually worth as a share of a voice minute.

What does Deepgram cost?

There is a free tier and self-serve signup, so every figure above is checkable in a browser without a sales call. In this directory 108 of 264 tools will not quote without one, which makes that worth more than it sounds.

This site's cost model takes $0.0077 rather than $0.0065, because the model states that it assumes list pricing. A buyer paying today pays $0.0065 and should knock $0.0012 off any figure here that includes speech recognition.

The strikethrough that is not an anchor

The page shows Flux English at $0.0065 with $0.0077 drawn through, under the heading Limited-time promotional rates on streaming, and publishes no end date.

The instinct is to read that struck-through figure as a fictitious anchor, the retail trick of inventing a higher price to make the real one look like a saving. This site's first record of Deepgram made exactly that assumption and stored $0.0065 as the list rate.

The archive says otherwise. On 10 April 2026 the same table carried $0.0077 as pay-as-you-go and $0.0065 for the committed Growth tier, in two separate columns, with no promotion running at all. The committed-tier price was later moved into the pay-as-you-go column and relabelled as a limited-time offer, and it has stood there since at least 18 May 2026.

So the anchor is real and the rate never fell. What changed is which column the cheaper number sits in, and what the page calls it.

A limited-time offer with no end is not limited

Set the two findings on this page side by side and the contrast is the story.

The Voice Agent increase carries a date, a percentage a buyer can compute, and enough notice to act. The streaming promotion carries a label implying urgency and no expiry anywhere on the page, and has stood for over three months. One of those tells a buyer what to do and the other tells them to hurry, for reasons that do not exist.

Neither is dishonest. They are simply two different jobs the same page is being asked to do, and the buyer's defence is the same in both cases: write the specific rate for your configuration into an order form rather than inheriting it from a page that can be edited without notice.

What the September increase actually costs

Two things land on 12 September 2026 rather than one. The Voice Agent rate rises, and the synthesis inside it stops being free. The second has no published replacement rate that this site holds, which means the table above is a floor on the increase rather than the whole of it.

A reseller sells Deepgram below Deepgram

Telnyx resells speech to text priced per model, and lists Deepgram Flux at $0.0074 a minute, read 16 August 2026.

Deepgram's own list for the same model is $0.0077. The reseller is 3.9% under the vendor [computed].

That is the opposite of how a supply chain is supposed to price, and it is a useful demonstration that the vendor's own page is not automatically the cheapest place to buy the vendor's own product. It is a small saving in absolute terms and the reason to notice it is structural: if a reseller can undercut list, list is a negotiating position rather than a floor.

Why Deepgram carries no all-in figure on this site

Six entities in this directory carry a computed all-in per-minute floor: Vapi, Retell, Bland, ElevenLabs Agents, Millis and Ultravox. Deepgram is not one of them.

For the streaming product the reason is straightforward: it is one of the four components, not a platform quoting an incomplete bundle, so there is nothing to add components to.

For the Voice Agent product the reason is a gap rather than a category. That rate bundles recognition and synthesis, and the page does not itemise what else it does or does not cover. Without a verified exclusion set there is no honest arithmetic to publish, so this site records not itemised on the page and leaves the cell blank rather than guessing.

An exclusion list is the input that decides everything downstream. It is also the field most likely to be filled in from a summary rather than from the page, which is how a directory ends up publishing a confident number derived from an assumption nobody checked.

Recognition is about a sixth of a voice minute

This site's component floor for a voice agent minute is $0.0465: recognition at $0.0077, a low-latency model at $0.0108, synthesis at $0.0140 and US telephony at $0.0140.

Recognition is 16.6% of it [computed], and the whole spread between the cheapest and dearest streaming vendors is measured in tenths of a cent. Synthesis, by contrast, runs from $0.014 to $0.085 per minute of call at list, a factor of 6.1.

So optimising your speech-to-text bill is the smallest lever on the board. The reason to choose carefully here is latency and turn detection, which decide whether the agent sounds like a person or a system, and this site treats that trade in the full costing.

Flux versus the cheaper models, and why the model uses Flux

Deepgram publishes several recognition models and Nova-3 is cheaper than Flux.

Flux is the conversational one, built for voice agents, with turn detection and interruption handling included. That is why this site's cost model takes it for a like-for-like comparison: a stack running a cheaper model without turn detection pays for that capability somewhere else, usually in an orchestration layer or in an engineer's month.

A rate card comparison that puts a cheaper model beside Flux and calls it a saving is comparing two different products.

Reliability, which the pricing page does not price

Deepgram publishes a machine-readable incident feed, which is more than most vendors in this category do.

Pulled as JSON on 25 August 2026 it carried 50 dated incidents, the most recent on 20 August 2026, elevated 5XX errors on the authentication grant endpoint. That is the current depth of the page rather than a lifetime total, and 50 incidents is not by itself a criticism: a vendor that publishes them is easier to evaluate than one that does not.

What it does tell you is that an authentication endpoint is a single point of failure for every call your agent is trying to start, and that belongs in your architecture review rather than in your pricing spreadsheet.

Free tiers and what they are actually for

Deepgram has a free tier and self-serve signup, which puts it in the minority of this directory: 108 of 264 tools here will not quote a price without a sales call.

A free tier on a recognition vendor is genuinely useful, because the thing you need to test is not the price. It is whether the model handles your callers' accents, your product names and your line quality, and that takes a corpus of your own audio rather than a demo clip.

What it will not establish is your unit cost at volume, because the interesting rates are the committed-use ones and none of those is published.

How to price your own recognition line

• Decide streaming or batch first. A live agent needs partial results while the caller is still speaking; a recorded-file workload does not, and the two are priced and engineered differently.
• Check whether turn detection is in the rate for the model you picked. If it is not, you are buying it again elsewhere.
• Use $0.0077 for modelling and $0.0065 for budgeting, and know which one you used. The promotion has no published end date, which is not the same as being permanent.
• Put 12 September 2026 in a calendar if you are on Voice Agent, because both the rate and the free synthesis change that day.
• Check a reseller's rate before assuming list is the floor. Telnyx lists the same Flux model at $0.0074.

What this page does not know

Every recognition rate this site can stand behind, with dates

Recognition is the component with the smallest spread and the most confusing rate cards, because the same vendor's model appears at four different numbers depending on who is selling it and in what unit. Everything below carries the day it was read.

The entire spread across the commercial rows is six tenths of a cent, $0.0058 to $0.0117, a factor of 2.0 [computed]. Set that against synthesis, which runs $0.014 to $0.085 per minute of call at list, a factor of 6.1, and the priority ordering writes itself.

Two rows are not comparable to the rest and are kept separate on purpose. The Voice Agent rows are per minute of call and bundle synthesis; every other commercial row is per minute of audio and bundles nothing. Multiplying the wrong one by your call volume is the arithmetic error this rate card invites.

When Deepgram is the wrong purchase

When you are buying Voice Agent as a like-for-like against a per-minute platform. It bundles recognition and synthesis and the page does not itemise what else it covers, so no all-in floor can be computed for it and no honest comparison against Vapi at $0.0965 or Retell at $0.0948 can be run. Get the exclusion list in writing or compare something else.

When you costed a deployment before 12 September 2026 and have not repriced it. Two things land that day: the Voice Agent rate rises 34%, and the synthesis inside it stops being free with no published replacement rate this site holds. At 20,000 minutes a month the rate change alone is $380 more per month and $4,560 a year [computed].

And when recognition is the line you are trying to optimise at all. It is 16.6% of a $0.0465 component floor [computed], and the whole vendor spread is tenths of a cent. Choose here on latency, turn detection and accuracy against your own audio, then go and spend the same afternoon on the voice, where the money actually moves.