Model

AssemblyAI holds its transcript price at $0.21 an hour as AWS spot GPUs cut its costs 70%

The $0.21-an-hour transcript now costs about 70% less to run than the same work on full-price compute, and the spread is bankrolling AssemblyAI's move into speaking and routing while rival Deepgram bought the open framework developers use to switch speech vendors.

Richard TangRichard Tangthree-minute read
Share
A person speaking close to the camera with live transcription text appearing under their mouth
AssemblyAI's realtime speech-to-text drawing live captions as someone speaks, in a clip the company posted to its own X account.

On November 16, AssemblyAI put its new transcription model, STT-3.7, into general availability and left the price at $0.21 an hour 1. Days earlier, in an AWS case study, the company said its speech-to-text inference now runs on EC2 Spot GPU instances, Amazon's−0.75% — Amazon, down 0.75 percent today reclaimable discount compute, cutting costs 70% against on-demand pricing while demand nearly quadrupled in a year 2.

Every voice agent in production buys a transcript, so input that costs far less to make at an unchanged price is the widest margin in the voice-agent stack. The saving is visible in what shipped around it: TTS-1, the company's first speaking model, in late October 3, and STT-3.7's Realtime-Agent variant, tuned for live agents rather than recordings 4.

The bill did not move, and the saving stayed with the vendor

STT-3.7 covers 32 languages, up from 18, and switches between them mid-sentence 1. A new diarization model makes 29% fewer speaker-attribution errors, the difference between a readable two-person call and a scrambled one 1.

The case study describes the mechanics that let bargain compute carry a production service: when Amazon reclaims a spot instance, a Lambda handler, one of AWS's event-run functions, returns in-flight jobs to a queue for retry 2. AssemblyAI's last announced round was a $50 million Series C led by Accel in December 2023, and it has stayed private, so this margin lands in its next funding round rather than on an exchange 9. The listed name in the chain is Amazon, which sells the capacity underneath 2.

The margin is buying the other side of the call

In a live agent the hard part is timing, which Deepgram's chief product officer, Ed Anuff, calls "mostly plumbing" 6. The Realtime-Agent model, at $0.45 an hour, targets that link with conversation memory, noise suppression and next-generation end-of-turn detection 5.

The speaking side arrived with TTS-1, at $15 per million characters, plus instant voice cloning and plugins for Pipecat and LiveKit, the open frameworks most voice agents are assembled in 38.

Listening costs $0.45 an hour and speaking about $0.81; Deepgram's whole agent is $4.50

Deepgram bundles the whole call; AssemblyAI sells the listening and speaking parts

  • Estimate
$0/hr$1/hr$2/hr$3/hr$4/hr$5/hrDeepgram Voice Agent API$4.5/hrIncludes the language-model stageAssemblyAI TTS-1 (speaking)$0.81/hrAssemblyAI STT-3.7 Realtime-Agent (listening)$0.45/hr
Data
Value
Deepgram Voice Agent API$4.5/hr
AssemblyAI TTS-1 (speaking) (estimate)$0.81/hr
AssemblyAI STT-3.7 Realtime-Agent (listening)$0.45/hr
List prices per hour of a live call. Speaking cost is OrcaRouter's estimate at about 54,000 characters an hour; Deepgram's rate bundles the language-model stage, AssemblyAI's parts leave the brain to the buyer. Sources: AssemblyAI pricing; Fullduplex; OrcaRouter.5,6,7

Its speaking voice lands fourteenth, 107 points back

In Artificial Analysis's Provider Voice Arena, where blind listening votes produce an Elo rating on the chess-style points scale, AssemblyAI-TTS-1 sits fourteenth at 1,227 7. Eleven v4, second at 1,334, leads it by 107 points, and Eleven v4 Turbo is first at 1,336 7. The model's weakest measured area is deciding how an abbreviation, address or account code should be spoken, with no pronunciation dictionary to configure around it; the rest of the gap is language count, voice inventory and controls 7.

AssemblyAI's new voice sits 107 Elo points behind the leader

Its voice is far cheaper per character and 107 points behind the best one

05001,0001,500Eleven v4 Turbo1,336Eleven v41,334AssemblyAI-TTS-11,227
Data
Value
Eleven v4 Turbo1,336
Eleven v41,334
AssemblyAI-TTS-11,227
Blind-listening Elo ratings in Artificial Analysis's Provider Voice Arena, as reported by OrcaRouter: Eleven v4 Turbo first, Eleven v4 second, AssemblyAI-TTS-1 fourteenth. Higher is better. Source: OrcaRouter.7

The rival bought the framework built for switching

On October 29, Deepgram, valued at $1.3 billion in its January funding round 6, said it was acquiring Daily, maker of Pipecat, the open-source framework that exists so developers can swap speech vendors (we reported it the day it was announced). The foundation meant to keep Pipecat neutral has no published charter 6.

Deepgram's own Voice Agent API, which bundles listening, thinking and speaking at $4.50 an hour, states in its settings that for speech-to-text "currently only Deepgram is supported" 6. AssemblyAI's answer is to sell the parts and stay swappable: its speaking model runs inside Pipecat as a plugin 8, and an LLM gateway beside it lists 60 models, from GPT to Claude to Gemini, so the thinking stage can sit on the same bill 5.

The gap gets priced at AssemblyAI's next round

The price likely held because nothing in the market forces it down: the nearest all-in product, Deepgram's bundle, sells for far more than a bare transcript 6. Deepgram acquired distribution; AssemblyAI built out speaking and routing with the saving it kept. Its next private round, marked against Deepgram's January valuation, is where that difference gets priced.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading AMZN

Reader comments

0 comments

    Model

    Z.ai falls to HK$335, more than 85% below its June peak, as Alibaba cuts token prices

    The Beijing lab still carries about HK$163 billion of market value on HK$1.49 billion of trailing sales, with a HK$12.4 billion annual cash burn underneath. A new US rule on rented compute and Alibaba's cut to token prices are squeezing what is left for the funds holding Hong Kong's AI listings.

    Model

    Anthropic's Thursday Debut Puts a Daily Price on Alphabet's Stake

    Alphabet holds the equity, rents Anthropic its compute and undercuts its prices, and from November 19 the market marks that bundle daily. The consensus says weaker American open source restores Anthropic's pricing power; the models actually holding the floor are Chinese.

    Model

    Meta's Q3 earnings fell to $3.83 against $5.46 expected, on a $9.89 billion teen-harm legal charge

    Excluding the charge Meta beat, and Businessweek's settlement pricing implies only about half the pending teen-harm claims are covered, with 2,000 school districts unpriced. The reserve note in January's filing decides.

    Model

    Nvidia shows how to run Z.AI's GLM-5.3 on $190,000 of Nvidia hardware, with no per-token bill to Z.AI

    The tutorial walks enterprises through serving GLM-5.3 on two DGX Stations, roughly $190,000 at listed reseller prices, with no per-token bill to Z.AI. For a stock carrying 118.6 times sales, that reads as a price list for everything Z.AI sells.

    Model

    Anthropic's Claude models tie at 58 on a coding index Cognition partly built, days before its Nasdaq IPO

    A third of the refreshed index runs on FrontierCode, a benchmark built by Cognition, whose Devin agent announced one of the two tied winners, Claude Fable 5.2, on the day of its release. Anthropic's model lists on Nasdaq on November 19; Cognition's Devin has offers from $70 billion on the table.

    Model

    Anthropic's Fable 5.2 passes 41.3% of real freelance jobs before its Nasdaq debut

    The one independent test of finished freelance work puts Anthropic's Fable 5.2 first at 41.3%, and the model still fails most of it, days before a Nasdaq debut priced at up to $1.4 trillion.

    Subscribe to TheInference for free