The cheapest speech-to-text API in 2026 is Fireworks running Whisper v3 large at $0.0015 per minute, followed by Speechmatics batch at $0.0022 and AssemblyAI Universal-2 at $0.0025. OpenAI's gpt-4o-transcribe, the model most people search for by name, is $0.006 per minute, roughly four times the cheapest option. Every price below is per minute of audio and is re-checked weekly.
| Provider | Model | Price per minute |
|---|---|---|
| Fireworks | Whisper v3 large | $0.0015 |
| Speechmatics | Batch | $0.0022 |
| AssemblyAI | Universal-2 | $0.0025 |
| OpenAI | gpt-4o-mini-transcribe | $0.0030 |
| ElevenLabs | Scribe | $0.0037 |
| OpenAI | gpt-4o-transcribe | $0.0060 |
| Deepgram | Nova-3 | $0.0077 |
| Gladia | Async Starter | $0.0102 |
| Whisper (self-host or hosted) | Whisper large v3 | varies by host |
How to read this table
Batch and async pricing is cheaper than realtime streaming at every provider, so confirm which mode a quoted rate covers before you budget against it. Accuracy also varies by accent, audio quality, and domain vocabulary, so the cheapest row is not automatically the right pick for a production workload. Prices move; each linked page carries the date of its most recent check.
Where the free credits are
Voice features (transcription and text-to-speech) are easy to add and easy to overpay for. Here is how to get speech AI for the lowest cost in 2026, starting with the providers that give you credits for free.
Start with free voice-AI credits
The fastest way to ship voice features for $0 is to use the signup credits these providers hand out:
- Deepgram: $200 in credits, usable across speech-to-text, text-to-speech and voice-agent APIs. The largest signup credit in the voice space.
- AssemblyAI: $50 in credits for transcription and speech understanding.
- ElevenLabs: a free monthly text-to-speech allowance (plus speech-to-text and voice design) on the free plan.
Those three alone cover a lot of an early product. See the always-current list in the catalog.
How to think about cost
Speech APIs are usually billed per minute (for STT) or per character (for TTS). When you compare, normalize to the same unit and watch for:
- Streaming vs batch: real-time streaming often costs more than pre-recorded batch transcription.
- Model tier: a smaller, faster model is frequently accurate enough and much cheaper.
- Free-tier limits: generous monthly allowances can cover low-volume apps indefinitely.
A low-cost voice stack
- Transcription: spend Deepgram's $200 credit first, then AssemblyAI's $50; both are strong and let you start with no card.
- Text-to-speech: ElevenLabs free tier for early volume, then compare per-character pricing as you scale.
- The LLM in the middle: keep it cheap by routing to the lowest-cost provider per model. See our rankings.
When you outgrow the free credits
Once you are past the credits, the same rule applies as with LLMs: the cheapest provider for the exact model and mode you need can be far below the default. Re-compare periodically, because voice pricing moves.
Grab the voice credits above with an account, and pair them with the rest of your AI stack from the catalog. Related: free AI API credits.