Is Grok Speech (STT + TTS APIs) Down?

THE `grok-voice-latest` REPOINT HAPPENED ON SCHEDULE 2026-08-05 -- AND IT IS A 1.6x PRICE INCREASE IF YOU DIDN'T PIN (verified against xAI's own models endpoint 2026-08-06): the alias migration announced on 7/29 executed as promised. xAI's live model catalogue now attaches **`aliases: ["grok-voice-latest"]` to `grok-voice-think-fast-2.0`** across every cluster it publishes (us-east-1, eu-west-1, us-saltlake-2), and **`grok-voice-think-fast-1.0` now carries no aliases at all**. So any integration still calling `grok-voice-latest` is on Think Fast 2.0 as of 8/05, with no code change and no deploy on your side. **THE PART XAI DID NOT PUT IN THE ANNOUNCEMENT: this alias move raised the per-audio-second rate by exactly 1.6x.** xAI's catalogue prices Think Fast 2.0's realtime audio at **1.6 times** Think Fast 1.0's (13,333,333 vs 8,333,333 in the endpoint's own units). Against the vendor-published **$0.08 per minute** headline for Think Fast 2.0, that puts Think Fast 1.0 at **$0.05 per minute** -- so unpinned callers went from roughly $3.00 to $4.80 per hour of audio overnight. The capability jump is real (AA Speech-to-Speech Quality Index 82.9% vs 75.7%, time-to-first-audio 0.70s vs 1.25s -- see the entry below), and 2.0 uses 0.4x the reasoning tokens of 1.0, so the effective cost story depends on your workload. But if you budgeted voice spend off the 1.0 rate and did not pin, **check your August bill** -- this is a silent 60% unit-price increase on a default alias. Pinning still works: `grok-voice-think-fast-1.0` remains callable, it simply is not what `latest` means any more

NEW MODEL -- GROK VOICE THINK FAST 2.0, AND `grok-voice-latest` REPOINTS ON 2026-08-05 (2026-07-29, vendor-primary): xAI shipped '**our most capable speech-to-speech voice model**', publishing third-party benchmarks from Artificial Analysis. **AA Speech-to-Speech Quality Index 82.9%** vs 75.7% for Think Fast 1.0, ahead of GPT-Realtime-2.1 High (79.1%) and Gemini 3.1 Flash High (69.5%). Component scores: Big Bench Audio 97.2%, **Full Duplex Bench 95.1%** (up from 77.8%, essentially closing the gap to GPT-Realtime-2.1's 95.7%), tau-voice Bench **56.5%** (vs 45.7% GPT-Realtime-2.1), and **time to first audio 0.70s** vs 1.25s for 1.0 and 2.98s for Gemini 3.1 Flash. Transcription: xAI claims a **1.5-2.0x accuracy improvement over Deepgram Nova 3 and ElevenLabs Scribe v2** across thousands of short phrases in 24 languages, widening to **~10x in noisy settings**. Architecture note worth understanding: these models **reason while speaking** -- 'reasoning in parallel with speech makes the model substantially smarter than other speech-to-speech models with no impact on latency' -- and 2.0 uses **0.4x the reasoning tokens** of 1.0, so tool calls typically fire before the agent finishes its first sentence. **PRICING: $0.08 per minute of audio.** **ACTION REQUIRED FOR API USERS: on August 5, 2026 the `grok-voice-latest` alias moves from `grok-voice-think-fast-1.0` to `grok-voice-think-fast-2.0`.** No action is needed to upgrade; to stay on 1.0 you must **pin `grok-voice-think-fast-1.0` before then**. xAI says the upgrade needs no prompt edits, and cites a Starlink support-line A/B test showing higher sales conversion and support containment. BRANDING NOTE: the vendor page is titled under **SpaceXAI** while the footer still reads (c) 2026 X.AI LLC -- the brand has shifted, the legal entity name has not

EXPANSION (2026-07-06): xAI released **21 new flagship TTS voices** (Lumen, Castor, Naksh, Atlas, Carina, Zagan, Helix, Orion, Luna, and more), bringing the lineup to 26 -- each cast for a specific job (support, characters, commentary, advertising, education) and natively multilingual across Grok Voice's 25+ languages. The original five (Ara, Eve, Leo, Rex, Sal) were retrained for more natural pacing/phrasing/emphasis. All are available in the realtime Voice Agent API, the Text-to-Speech API, and a **new no-code Grok Voice Agent Builder** in the xAI console. **Custom voice cloning from ~1 minute of audio is now exposed** (closes a prior gap vs ElevenLabs). Speech tags ([pause], <whisper>, <emphasis>, <soft>) control delivery

Launched 2026-04-17 -- expect first-week rate-limit surprises and occasional multilingual hiccups in the streaming path. xAI's console rate limits are documented but may be adjusted during the shakedown period

The xAI-to-SpaceX acquisition (2026-02-02) means billing, compliance, and procurement flow through SpaceX. For US-regulated customers (healthcare, finance, defense) the vendor-approval pathway is new and may take longer than with a standalone AI vendor

Speech tag control is powerful but underdocumented at launch -- expect trial-and-error to dial in the right [laugh]/[whisper]/<pause> mix for conversational TTS