Deepgram rates applied to a full-mesh huddle, where every participant is a separately billed audio stream.
CallFX runs a full mesh. Audio travels peer to peer and never reaches a server, so there is no mixed recording anywhere to transcribe once. Each participant's browser transcribes its own microphone and broadcasts the text over the data channel the huddle chat already uses.
That makes speaker labels free and correct — every stream is one known person — but it means a four-person call costs four times a one-person call of the same length. Every estimate below is in stream-minutes, not call-minutes.
Server-side transcription would bill 45 minutes instead of 180 — but getting there means routing audio through an SFU or a headless recorder bot that joins the call. That is new infrastructure to build, host and keep alive, and it re-introduces the speaker-labelling problem the mesh solves for free.
At these rates the mesh premium on one huddle is about a quarter of a dollar. It stops being obviously worth it somewhere north of a thousand huddle-hours a month.
| Model | Pay-as-you-go | Growth | $ / stream-hour |
|---|---|---|---|
| Nova-3 MultilingualHinglish | $0.0058 | $0.0050 | $0.348 |
| Nova-3 Monolingual | $0.0048 | $0.0042 | $0.288 |
| Flux Multilingual | $0.0078 | $0.0068 | $0.468 |
| Flux English | $0.0065 | $0.0057 | $0.390 |
Nova-3 Multilingual is the row that matters. Its multi mode code-switches without being told which language is being spoken, and Hindi is in that set alongside English — so a student who slides between the two mid-sentence still gets a usable transcript. Flux is faster but its multilingual model covers only ten languages and costs more.
| Model | Pay-as-you-go | Growth |
|---|---|---|
| Nova-3 Multilingual | $0.0052 | $0.0043 |
| Nova-3 Monolingual | $0.0043 | $0.0036 |
| Whisper Large | $0.0048 | $0.0048 |
| Huddles / month | Stream-hours | Transcription | Total incl. summaries |
|---|
Deepgram bills the audio you stream. In a four-way conversation each person is quiet most of the time, so closing the socket during silence is the single largest saving available — and it costs one VAD check in the browser.
Monolingual Nova-3 is $0.0010/min cheaper. Start a huddle on multi, and if nothing but English appears in the first minute, reconnect on the monolingual model.
Annual pre-paid credits, drawn down against real usage. Worth it once monthly spend is predictable — not before, since unused credits expire.
Most huddles do not need a transcript. Gating it behind the existing recording-consent flow cuts spend and settles the FERPA question in the same move.
| Provider · model | $ / stream-hour | Billed on | Hindi code-switching |
|---|---|---|---|
| Deepgram Nova-3 Multilingual | $0.348 | Audio streamed | Documented |
| Deepgram Nova-3 Monolingual | $0.288 | Audio streamed | No |
| AssemblyAI Universal-Streaming | $0.150 | Socket open time | Unconfirmed |
| AssemblyAI Universal-3.5 Pro Realtime | $0.450 | Socket open time | Unconfirmed |
The headline gap is real but narrower than it looks, for two reasons.
The billing units differ. AssemblyAI charges for how long the WebSocket stays open, not how much audio crosses it. In a huddle the socket is open for the whole call whether or not anyone is speaking — which is exactly the saving the VAD lever depends on, and it does not apply there.
The language answer is missing. AssemblyAI's streaming documentation does not publish a language list, and its pricing page's "100+ languages" refers to translation add-ons rather than recognition. Until that is confirmed in writing, it cannot be treated as a Hinglish-capable option — which is the one requirement that decides this for an Indian student product.