Deepgram

What Deepgram actually meters

Four products, three different units, and the unit changes with the endpoint rather than the account.

Speech to text meters seconds of audio. Deepgram is unusually direct about this being a competitive claim rather than an implementation detail, arguing in its own FAQ that rivals rounding to 15 seconds or a full minute inflate invoices by 15 to 20%. On short clips, which is most voice agent traffic, that difference compounds.

Text to speech meters characters of input, not seconds of output, so the bill is set when you write the prompt rather than when the audio plays.

Voice Agent meters conversation duration, covering the transcription, the model call and the synthesis as one orchestrated minute. Bring your own model and you pay only for the audio parts, which is why the Custom tier at $0.050 undercuts Standard at $0.075.

Audio Intelligence bills on top. Summarisation, topic detection and sentiment analysis each carry their own per-hour or per-1,000-token charge, added to the transcription cost rather than included in it.

One exception undoes the per-second generosity. Multichannel audio bills total processed duration, so a 10 minute stereo file costs 20 minutes. Anyone transcribing two-sided call recordings should model that before comparing rates.

How credits work

Both plans draw down a prepaid balance rather than invoicing in arrears.

Pay As You Go opens with $200 of credit, no card required, which Deepgram puts at roughly 43,000 minutes on Nova. It carries no expiry date, so it sits until spent rather than lapsing after a year.

Growth converts that into an annual commitment from $4,000, buying higher concurrency and priority support. The plan card badges "Save up to 20%", but Deepgram's own FAQ puts Growth rates "about 12.5% lower" and the rate card agrees: Nova-3 streaming drops from $0.0048 to $0.0042, while Voice Agent Standard drops from $0.075 to $0.068, a 9.3% cut. Auto-load defaults to adding $100 whenever the balance falls to $10, and can be turned off.

Refunds are narrow and clearly stated: unused credits bought in the last 30 days only. Anything older is non-refundable on the reasoning that the service was available whether or not you used it.

What happens when you hit the limit

Two different failures, depending on what you exhausted.

Run out of credits and Growth customers do not fall back to list price. Overage bills at your Growth rate plus a 10% premium plus tax, and it settles every Monday for the previous week rather than at month end. Deepgram documents an escape: buy a new Growth plan early, which resets your credits at the standard rate and rolls the billing cycle forward.

Run out of concurrency and requests queue or are rejected outright. Pay As You Go caps text to speech and Voice Agent at 45 concurrent connections against Growth's 60, with Audio Intelligence at 10 on both.