What is the best speech-to-text API in 2026?

For most production audio as of September 2026: SpaceXAI's Grok Voice Transcribe 2.0. It ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard and is the cheapest major API at $0.10 per hour batch and $0.20 per hour streaming, with diarization and timestamps included. AssemblyAI Universal-3.5 Pro is the pick when you need audio-intelligence add-ons or medical mode, ElevenLabs Scribe v2 for 90+ languages, and self-hosted Whisper large-v3 when audio cannot leave your machines.

How much does speech-to-text cost per hour in 2026?

List prices per hour of audio, batch, verified September 20, 2026: Grok Voice Transcribe 2.0 $0.10; Meta Muse Voice Transcribe $0.18; AssemblyAI Universal-3.5 Pro $0.21 (+$0.02 diarization); ElevenLabs Scribe v2 $0.22; Deepgram Nova-3 $0.26; OpenAI GPT Transcribe $0.27; OpenAI Whisper API $0.36; self-hosted Whisper $0 plus compute. At 10,000 hours a month that is $1,000 on Grok versus $3,600 on the Whisper API.

Is Whisper still worth using in 2026?

Self-hosted, yes: Whisper large-v3 and forks such as faster-whisper remain the default when privacy or offline use matters, and the hosted alternatives publish no open weights. As a paid API, no: OpenAI's own GPT Transcribe is cheaper at $0.0045 per minute, and every provider on this list beats the $0.006-per-minute Whisper API on price while matching or exceeding it on accuracy for noisy, accented or multi-speaker audio.

Which speech-to-text API is best for voice agents?

Streaming price and turn detection decide it. Grok Voice Transcribe 2.0 streams at $0.20 per hour with smart turn detection and led SpaceXAI's telephony tests. Deepgram Nova-3 streams at $0.29 per hour with the most mature voice-agent ecosystem. AssemblyAI's Voice Agent API bundles STT, LLM and TTS on one WebSocket at a flat $4.50 per hour. ElevenLabs Scribe v2 Realtime has ~150 ms latency at $0.39 per hour and pairs naturally with ElevenLabs TTS.

Quick Answer

Best Speech-to-Text APIs 2026: Top 8 Ranked by Price

Published: September 20, 2026

The ranking (September 2026)

#	API	Batch $/hr	Streaming $/hr	Standout	Open weights
1	Grok Voice Transcribe 2.0 (SpaceXAI)	$0.10	$0.20	#1 of 32 streaming models on Artificial Analysis; diarization, timestamps, key terms included	No
2	AssemblyAI Universal-3.5 Pro	$0.21	$0.45	Code-switching in 18 languages, entity accuracy, Medical Mode, Sync API	No
3	ElevenLabs Scribe v2	$0.22	$0.39	90+ languages, 48-speaker diarization, non-speech events, ~150 ms realtime	No
4	Meta Muse Voice Transcribe	$0.18	$0.18	Same price both modes, 20+ speaker realtime diarization, zero-retention option	No
5	Deepgram Nova-3	$0.26	$0.29	Most mature streaming/voice-agent stack, custom models, on-prem	No
6	OpenAI GPT Transcribe	$0.27	via Realtime API	Tight integration with OpenAI agent workflows	No
7	Whisper large-v3 (self-hosted)	$0 + compute	—	Only open-weight option in this list; runs offline	Yes
8	Microsoft MAI-Transcribe-1	Azure pricing	Azure	Best inside Teams, Copilot and Azure AI Foundry	No

Verified September 20, 2026. Prices are pay-as-you-go list rates per hour of audio; annual and volume tiers are lower everywhere.

1. Grok Voice Transcribe 2.0 — best overall

Released September 18, 2026, Transcribe 2.0 roughly doubled the accuracy of SpaceXAI’s 1.0 model at an unchanged $0.10/hr batch, $0.20/hr streaming. It ranks first for accuracy among 32 streaming models on Artificial Analysis’s public leaderboard, and on SpaceXAI’s production sets short multilingual phrases dropped from 20.6% to 6.8% WER and telephony first-final transcripts from 3.9% to 2.7%. Diarization, word-level timestamps with confidence, up to 100 key terms per request, 8-channel transcription, filler removal and smart turn detection are all included. API-only at api.x.ai/v1/stt; Atlassian Loom is the launch customer. Full comparison: Grok Transcribe 2.0 vs Deepgram vs AssemblyAI vs Scribe v2.

Pick it for: call centres, in-car and IVR audio, any high-volume workload where the bill matters.

2. AssemblyAI Universal-3.5 Pro — best feature stack

$0.21/hr async, $0.45/hr realtime, billed per second. Diarization (+$0.02/hr async), keyterm prompting and Medical Mode (+$0.15/hr) are add-ons. AssemblyAI’s differentiators are the layers on top of transcription — summaries, topic detection, PII redaction, entity accuracy — and the July 14, 2026 Sync API for short clips. Its own comparisons claim wins over Scribe v2 on code-switching and entity error rate.

Pick it for: healthcare, legal, compliance, and pipelines that need structured outputs from audio.

3. ElevenLabs Scribe v2 — best multilingual batch

$0.22/hr batch, $0.39/hr realtime ($0.28 on annual Business plans). 90+ languages, diarization up to 48 speakers, non-speech event tagging, keyterm prompting (+$0.05/hr), entity detection (+$0.07/hr). Reports about 2.3% AA-WER on the non-streaming index. Natural pairing with ElevenLabs TTS for dubbing and voice agents.

Pick it for: podcasts, video localisation, media libraries in many languages.

4. Meta Muse Voice Transcribe — simplest pricing

Launched September 1, 2026 at a flat $0.18/hr for both batch and streaming, billed per second rounded down, with realtime diarization for 20+ speakers included and a zero-data-retention option at the same price. It was the price leader for 17 days until Grok 2.0. Independent accuracy data was thin as of September 20.

Pick it for: teams standardising on Meta’s Muse API who want one price and no add-on maths.

5. Deepgram Nova-3 — best voice-agent ecosystem

$0.26/hr batch, $0.29/hr streaming, verified September 8, 2026, with $200 free credit. Sub-300 ms streaming, custom vocabulary and model training, on-prem deployment. Most voice-agent frameworks integrated Deepgram first, and that ecosystem is still its strongest argument now that it is undercut 2.6x on batch price.

Pick it for: existing voice agents, custom domain models, enterprises that need on-prem.

6. OpenAI GPT Transcribe — best inside OpenAI workflows

$0.0045/min ($0.27/hr), OpenAI’s recommended file-transcription model; a diarization variant and the Realtime API cover streaming. Choose it when transcripts feed directly into GPT-5.6 or GPT-6 Astra agents and you want one vendor.

7. Whisper large-v3 — best when audio cannot leave your machines

Still the only serious open-weight option in this list. Self-hosted with faster-whisper or MLX it runs 10–15x real-time on an M4 Max and costs nothing per hour. As a hosted API ($0.006/min, $0.36/hr) it is now the most expensive and least accurate on noisy audio — use GPT Transcribe or any provider above instead. Guide: Whisper vs MAI-Transcribe-1 vs Deepgram.

8. Microsoft MAI-Transcribe-1 — best in the Microsoft stack

Powers Teams transcription and Copilot Voice; available through Azure AI Foundry. Choose it when procurement, identity and data residency are already Azure. See what is MAI-Transcribe-1.

Cost at scale

Monthly bill for 10,000 hours of batch audio at list price: Grok $1,000 · Muse $1,800 · AssemblyAI $2,100 (+$200 diarization) · Scribe v2 $2,200 · Deepgram $2,600 · GPT Transcribe $2,700 · Whisper API $3,600 · self-hosted Whisper ≈ GPU rental only. Streaming at the same volume: Grok $2,000 · Muse $1,800 · Deepgram $2,900 · Scribe v2 $3,900 · AssemblyAI $4,500.

How to choose in one minute

Audio must stay on-prem? Whisper self-hosted. Nothing else gives you weights.
Cheapest accurate transcripts? Grok Voice Transcribe 2.0.
Need summaries, PII redaction, medical vocabulary? AssemblyAI.
Many languages, media content? ElevenLabs Scribe v2 — benchmark against Grok on a 10-file sample.
Already on Deepgram or Azure? Stay unless volume makes the price gap material.