Deepgram Pricing 2026

Deepgram offers a Pay-As-You-Go plan with $200 in free credits (no expiration) followed by usage-based pricing. Nova-3 Monolingual costs $0.0077/min ($0.46/hr), Nova-3 Multilingual costs $0.0092/min ($0.55/hr), and the Base model costs $0.0145/min ($0.87/hr). Growth plan prepays $4,000+ annually for up to 20% discounts (Nova-3 at $0.0065/min). Voice Agent API costs $0.08/min for standard agents. Enterprise plans start at $15,000+ annually with custom pricing, dedicated support, and optional self-hosted deployment. Billing is per-second with no rounding up to full minutes.

Deepgram offers $200 in free credits with no credit card required and no expiration, making it one of the most generous free tiers among transcription APIs. At $0.0077/min for Nova-3, this covers approximately 26,000 minutes (~433 hours) of transcription. Unlike AssemblyAI's one-time $50 credit or OpenAI Whisper's lack of free tier, Deepgram's $200 credit never expires and can be used across all services (speech-to-text, text-to-speech, Voice Agent API). For long-term free usage, Speechmatics offers 480 minutes/month recurring, but Deepgram's $200 upfront is more generous for initial testing.

Deepgram is an AI-powered speech recognition platform providing speech-to-text, text-to-speech, and Voice Agent APIs for developers. It specializes in real-time streaming transcription with ultra-low latency (<300ms) and per-second billing precision. Deepgram's Nova-3 model is optimized for accuracy, speed, and cost, supporting 30+ languages. The platform is used by companies like NASA, Spotify, and Twilio to power live captioning, voice agents, meeting transcription, and customer call analytics. Deepgram differentiates itself through streaming performance, transparent per-second billing, and $200 in credits with no expiration.

Deepgram Nova-3 costs $0.0077/min ($0.46/hr) vs AssemblyAI Universal at $0.0025/min ($0.15/hr), making AssemblyAI 3x cheaper per hour for basic transcription. However, Deepgram's per-second billing (vs per-hour) means you pay only for audio duration, not rounded-up hours. Deepgram's $200 free credit (no expiration, ~433 hours) is 4x more generous than AssemblyAI's $50 credit (~185 hours). Deepgram excels at real-time streaming with <300ms latency, while AssemblyAI offers richer audio intelligence features (entity detection, topic detection, auto chapters) bundled into one API. Choose Deepgram for real-time applications and latency-critical use cases; choose AssemblyAI for batch processing with advanced audio intelligence.

Deepgram Nova-3 at $0.0077/min is 62% more expensive than Rev AI's Reverb Turbo at $0.10/hr ($0.00167/min) but offers real-time streaming with <300ms latency that Rev AI cannot match. Rev AI's Reverb at $0.20/hr ($0.00333/min) is 2.3x cheaper than Deepgram for standard batch transcription. Deepgram bills per-second while Rev AI rounds up to the nearest second with a 15-second minimum. Deepgram's $200 free credit covers ~433 hours vs Rev AI's lack of a free tier. Choose Deepgram for real-time streaming, live captioning, and voice agents; choose Rev AI for budget-conscious batch transcription or when human transcription fallback ($1.99/min) is required.

Deepgram's base pricing includes speech-to-text transcription with automatic punctuation, capitalization, and per-second billing precision. Nova-3 Monolingual ($0.0077/min) supports English and common languages, while Nova-3 Multilingual ($0.0092/min) supports 30+ languages. Real-time streaming with <300ms latency is included at the same per-minute rate. Voice Agent API ($0.08/min) includes transcription, turn-taking, and conversational logic for building voice assistants. Text-to-speech via Aura-2 costs $0.030 per 1,000 characters (~$1.80/hr of generated audio). Audio Intelligence features are priced separately at $0.0003-0.0006 per token. Growth plans add 20% discounts with $4,000+ prepayment. Enterprise includes custom models, self-hosting, and dedicated support.

Yes, Deepgram charges for the full duration of submitted audio files, including silence, music, and non-speech segments, based on audio file length. However, Deepgram's per-second billing precision (unlike competitors that round up to full minutes or hours) means you pay exactly for the audio duration submitted. For real-time streaming, you are charged for the WebSocket connection time, not audio duration. To minimize costs, preprocess audio to remove long silences using voice activity detection (VAD) or FFmpeg before sending to Deepgram's API, or use Deepgram's built-in features to optimize streaming connections.

Deepgram operates on a usage-based billing model with no subscriptions or advance payments for Pay-As-You-Go customers, so there are no refunds -- you are billed only for audio processed. The $200 free credit is non-refundable and does not expire. Growth customers with $4,000+ prepaid commitments should negotiate refund terms directly in their contracts, as prepaid credits typically expire annually (12 months) and are non-refundable for unused balances. If you encounter a service issue or are overcharged due to a bug, contact Deepgram support to request a credit adjustment.

Deepgram's $200 free credit with no expiration is one of the most generous free tiers, covering approximately 26,000 minutes (~433 hours) of Nova-3 transcription at $0.0077/min. However, once this credit is exhausted, you move to Pay-As-You-Go pricing with no free monthly refresh. For ongoing free usage, Speechmatics offers 480 minutes/month recurring (5,760 minutes/year), which refreshes monthly. Deepgram's $200 upfront is better for initial testing and prototyping, while Speechmatics is better for sustained low-volume usage. For production applications, Deepgram's $0.0077/min is cost-effective for real-time streaming and high-concurrency workloads.