xAI's speech-to-text API is generally available and paired with text-to-speech

What we checked

• When did the xAI speech-to-text API page first appear or materially change? partly confirmed. The STT API was marked generally available in April 2026, an STT playground shipped 3 August 2026, Transcribe 2.0 became the default in September, and the console renamed the product Grok Voice Transcribe on 11 September 2026. Wayback CDX shows only one 200 snapshot of the marketing URL, on 11 October 2026, already branded Transcribe 2.0, so the URL's first-publish date remains unknown. [1][12][13][14][2]
• Is there a launch note, changelog, API reference, or model card confirming the voice API surface? confirmed. Yes. xAI published a joint STT/TTS launch post, developer release notes, a Speech to Text guide, REST and streaming API reference, a grok-voice-transcribe-2.0 model page, console changelog entries, and a public pricing table that all describe the same /v1/stt and /v1/tts surface. [7][1][4][5][8][6]
• Is speech-to-text the companion half of a material voice API that also includes documented text-to-speech, with pricing and limits? confirmed. TTS was marked generally available in March 2026, before STT. Docs specify POST /v1/tts and wss://api.x.ai/v1/tts at $15.00 per 1 million characters, plus speech-to-speech at $0.08 per minute. The STT product page sits next to a matching TTS marketing page. [1][9][6][15][11]
• Do the speech-to-text marketing page's capability claims match current API documentation? partly confirmed. Core surface matches: REST plus WebSocket, grok-voice-transcribe-2.0, timestamps, diarization, inverse text normalization, and keyterm boosting. Language count (25+ versus 38+), channel count (two versus eight), and a 20-speaker diarization cap do not match the API guide. [3][4][5]

What we found

• Developer release notes mark Text-to-Speech generally available in March 2026 and Speech to Text generally available in April 2026, initially with 25 languages and batch plus streaming modes. MarkTechPost independently dated the joint launch 18 April 2026. A September release-note item added grok-voice-transcribe-2.0 as the default. An October item says grok-voice-transcribe-1.0 reached end of life on 2 October 2026 and is routed to 2.0 at the same price. [1][2]
• The marketing page and Speech to Text guide document POST https://api.x.ai/v1/stt for batch (multipart file or URL, 500 MB cap, option fields before the file) and wss://api.x.ai/v1/stt for streaming. Clients must wait for transcript.created before sending raw audio. The default model is grok-voice-transcribe-2.0. Documented features include inverse text normalization, keyterm boosting (100 terms, 50 characters each), word-level timestamps, diarization, filler-word stripping by default, a VAD threshold, and Smart Turn on streaming. [3][4][5]
• Official pricing lists STT at $0.10 per hour for REST and $0.20 per hour for streaming, matching the launch blog. TTS is $15.00 per 1 million characters. Speech to Speech is $0.08 per minute for grok-voice-3.0, with grok-voice-2.0 deprecated at the same audio rate. The grok-voice-transcribe-2.0 model page exists, but pricing and rate-limit fields did not render in this session's fetch, so concurrency numbers are not cited. [6][7][8]
• TTS is a documented companion: POST /v1/tts (60,000-character unary cap) and wss://api.x.ai/v1/tts, default voice eve, speech tags, telephony codecs, and GET /v1/tts/voices. Voice Overview also documents realtime speech-to-speech, custom voices from clips up to 120 seconds, Zero Data Retention, and SOC 2 Type II, HIPAA-eligible, and GDPR claims. Console changelog added an STT playground on 3 August 2026 and renamed the product Grok Voice Transcribe on 11 September 2026. [9][10][11][12]
• Marketing copy and API docs diverge. The product page says 25+ languages and up to two channels; the STT guide lists 38+ languages and up to eight channels, with Opus streaming mono only. The product page claims diarization for up to 20 speakers; the API reference documents a speaker integer without that cap. Landing-page competitor charts did not yield extractable axis numbers. Implement against the API reference, not the landing page. [3][4][5]
• The launch post publishes vendor word-error rates of 5.0% on phone-call entities, 2.4% on video/podcasts, 10.9% on meetings, 9.3% on telephone, and 6.9% overall. MarkTechPost repeated those figures as xAI claims; no independent benchmark paper was found. Wayback Machine CDX lists a single HTTP 200 capture of the marketing URL on 11 October 2026, already advertising grok-voice-transcribe-2.0 and matching the live page. [7][2][13][14]

What it means

Treat Grok Voice Transcribe as a production API. Pin grok-voice-transcribe-2.0, send option fields before the file part, wait for transcript.created before streaming audio, and proxy WebSockets so keys never reach the browser. Use REST for files and streaming plus Smart Turn for agents. Budget $0.10 per hour batch and $0.20 per hour streaming for STT, and $15 per million characters for TTS. Do not use MarkTechPost's launch-era $4.20 per million TTS figure in a live cost model.

For full-duplex agents with tools, prefer speech-to-speech on grok-voice-3.0 or grok-voice-latest instead of chaining STT to a chat model to TTS. Keep standalone STT and TTS for captions, recordings, and telephony. Confirm Zero Data Retention, BAA, and residency in console and contract language before sending regulated audio; those claims appear on Voice Overview and were not independently audited here.

Still open

• When was https://x.ai/voice/speech-to-text first published? Wayback CDX shows only one 200 snapshot, dated 11 October 2026.
• Do independent benchmarks reproduce xAI's word-error-rate table against ElevenLabs, Deepgram, and AssemblyAI?
• Did TTS list price change after launch? MarkTechPost reported $4.20 per 1 million characters on 18 April 2026; current docs list $15.00. The launch-post extract obtained here did not include a TTS dollar figure.
• Is there a documented hard cap of 20 diarized speakers, or is that marketing-only copy?