xAI Just Dropped Grok STT 1.0 and It's Free on OpenRouter. Whisper's Last Stand Is Over.

Hey guys, Mr. Technology here.

It is Monday, July 27, 2026, and the most consequential model launch of last week was not Claude Opus 5. Don't get me wrong — Opus 5 is the Anthropic story. But Opus 5 is a frontier-text refinement. It is a workhorse model priced to win the default-model vote. It is the kind of release that shows up on a quarterly earnings call.

Grok STT 1.0 is the kind of release that changes an industry. xAI shipped its first dedicated speech-to-text model on July 23, 2026, the same week Anthropic finished its Claude 5 lineup and OpenAI was busy plugging its AI keypad into your desk. The headline is not the model. The headline is the price: free on OpenRouter as of July 24, with the xAI native API endpoint at api.x.ai/v1/stt and a streaming WebSocket at wss://api.x.ai/v1/stt. Every other speech-to-text vendor on Earth — OpenAI Whisper, ElevenLabs Scribe, Deepgram Nova-3, AssemblyAI Universal-2, Google Chirp 2, Microsoft Azure Speech — has spent the last 96 hours repricing.

This is the speech-to-text OpenAI Moment, except it is xAI swinging the axe and OpenAI is not even on the list. Let me give you the spec, the strategy, and the take.

What Shipped

• Endpoint: POST https://api.x.ai/v1/stt (multipart form upload, file max 500 MB), plus wss://api.x.ai/v1/stt for streaming.
• Languages: 26 explicitly supported for formatting — English, Arabic, Czech, Danish, Dutch, Filipino, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Macedonian, Malay, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Thai, Turkish, Vietnamese. The model transcribes any of these without the language parameter; setting it just enables Inverse Text Normalization (number/currency/unit formatting).
• Audio formats: 9 container formats auto-detected (WAV, MP3, OGG, Opus, FLAC, AAC, MP4, M4A, MKV) plus 3 raw formats (PCM, µ-law, A-law) with explicit audio_format and sample_rate.
• Channel support: mono, stereo, or up to 8 channels with multichannel=true — the response returns per-channel transcripts in the channels array.
• Speaker diarization: diarize=true adds an integer speaker field to every word.
• Keyterm biasing: up to 100 keyterms, each up to 50 characters, passed via repeated keyterm fields. This is the single biggest feature for product teams shipping voice agents on top of domain vocab.
• Word-level timestamps: every word gets a start and end float in seconds. The response shape is { text, language, duration, words[] } with no separate alignment call.
• Streaming: WebSocket binary frame ingest, interim results every ~500 ms via interim_results=true, configurable endpointing (silence-detection) from 0 ms to 5000 ms, default 10 ms.
• VAD threshold: tunable vad_threshold 0.0–1.0 (default 0.5). Lower values transcribe quiet / narrowband / telephony audio; 0 disables the gate entirely.
• Filler word handling: filler_words=true keeps uh/um/er in both the transcript text and the words array. Default false strips them — which is what every consumer voice-agent product actually wants.
• Inverse Text Normalization: format=true with a language converts "one hundred dollars" → "$100", "twenty three degrees" → "23°", "five miles per hour" → "5 mph". This is the boring feature that destroys every TTS→STT glue script in your codebase.

That is a real product spec. It is not a Whisper wrapper. It is not a "we finetuned Whisper on our data" gesture. The eight-channel multichannel, the 100-keyterm biasing, the streaming WebSocket with interim results, the VAD threshold, the ITN pipeline — none of that ships in OpenAI's whisper-1 and only some of it ships in Deepgram's Nova-3.

The Price Is the Story

OpenRouter lists Grok STT 1.0 at free as of the July 24 catalog update. xAI's native pricing for the synchronous endpoint has not been publicly broken out by per-minute rate, but the OpenRouter free tier means the marginal cost to a developer integrating this into a voice-agent prototype today is zero. That is not a pricing strategy. That is a market-entry strategy, and it is the same playbook xAI ran on Grok 4.5 in early July — drop the price to a different tier and force the rest of the field to either match or admit they are overcharging.

For comparison, the current STT pricing stack as of July 26:

• OpenAI gpt-4o-transcribe: $0.006 / minute input audio (the new flagship, replacing whisper-1)
• OpenAI whisper-1: $0.006 / minute
• Deepgram Nova-3: $0.0043 / minute on the pay-as-you-go tier, with self-serve at $0.0029
• AssemblyAI Universal-2: $0.0065 / minute
• ElevenLabs Scribe v1: $0.0066 / minute
• Google Cloud Chirp 2: $0.004 / minute on the standard tier
• Azure Speech: $0.006 / minute for the real-time tier
• xAI Grok STT 1.0 (OpenRouter free tier): $0.00 / minute

If you are a voice-agent startup building on LiveKit, Pipecat, or Vapi and your STT bill is non-trivial — say, 50,000 minutes of audio per month at $0.006 — your monthly bill just dropped from $300 to $0 with the same diarization, same keyterm biasing, same WebSocket streaming, same word-level timestamps, and same language coverage you were getting from OpenAI. That is not a 50% discount. That is the entire line item. The savings go straight to runway. For a Series A voice-agent company burning $200K per month, this is a one-line CFO email.

What xAI Gets Out of This

xAI does not lose money on the free tier in the way OpenAI loses money on free ChatGPT. STT inference is cheap — a 30-second clip costs roughly $0.0003 of compute on Grok STT 1.0 against an H100, which is the same order of magnitude as the GPU memory bandwidth cost of just loading the model. xAI is paying Colossus to run this. The free tier is a marketing cost, not a margin loss.

What xAI gets in return is a hard dependency. Once a voice-agent startup wires Grok STT 1.0 into their production pipeline — once they build their VAD, their keyterm lists, their diarization-aware post-processing on top of xAI's exact response shape — switching costs become real. The 100-keyterm limit, the 8-channel ceiling, the WebSocket protocol, the 500 MB upload cap, the ITN pipeline — these are the operational footprint of Grok STT in your codebase. The day xAI introduces pricing, the day a competitor matches their feature set, the day someone ships an open-weights alternative — your switching cost is the integration debt.

This is the same play Mistral ran with their API in 2024 and 2025, except at zero list price. Free is the new lock-in.

The second thing xAI gets is data. The 26 languages, the 8 channels, the diarization, the keyterm biasing — every audio file you send to the free tier becomes a training signal in some future Grok STT 1.5 or 2.0. OpenAI did this with Whisper for three years and it built the world's largest STT training set. xAI is starting the clock today.

The third thing xAI gets is Grok Voice. The xAI voice-agent platform has been in private beta since late June. The STT API is the front half of the voice agent stack; Grok Voice is the back half. Shipping STT for free, on a multi-channel streaming endpoint, with diarization and keyterm biasing, is the prerequisite for shipping voice agents at scale. You cannot be a serious voice-agent vendor in 2026 without a first-party STT. xAI now has one. OpenAI has one. Anthropic has nothing. Google has one but does not sell it as a standalone product at this price point. Deepgram is now structurally exposed.

What This Does To The Voice-Agent Stack

Voice agents are the fastest-growing software category of 2026. The stack is: STT → LLM → TTS, with VAD and turn-taking glue. Until this week, the dominant production choice was OpenAI Whisper or Deepgram on the STT side, GPT-4o or Claude on the LLM side, ElevenLabs or Cartesia on the TTS side. Every voice-agent framework — Vapi, LiveKit Agents, Pipecat, Retell, Bland, PolyAI, Slam — has STT provider abstraction because the providers are not interchangeable on price, feature, or response shape.

Grok STT 1.0 changes two things in that stack:

1. STT becomes a free commodity. The vendor that was charging $0.006/min for the same feature set is now competing with $0/min. The pricing floor in this segment just got reset. 2. xAI becomes a single-vendor voice-agent shop. You can now run Grok STT 1.0 → Grok 4.5 → Grok Voice on a single-vendor bill. The integration story is cleaner, the latency story is better (same network, same auth, same SDK), and the lock-in story is what xAI is buying.

For the open-source voice-agent frameworks (Pipecat, LiveKit Agents, the new MemGPT stack), the Grok STT 1.0 release is unambiguously good. They get a free, high-quality STT provider to add to their model registry. They get a marketing story ("built on Grok STT") that the Vapi-and-Bland tier cannot match without cutting their margins. They get a forcing function on Deepgram and AssemblyAI to either price-match or lean harder on enterprise features (custom vocab, on-prem deployment, compliance).

For ElevenLabs, which has been building a vertically-integrated voice-agent platform (STT + LLM orchestration + TTS), this is a direct hit on the "one throat to clear" story. ElevenLabs Scribe v1 is a fine model but it is now competing against free on the STT half of its platform.

For Deepgram, this is the harder problem. Deepgram has been the high-end STT vendor — best-in-class accuracy, lowest latency, real-time streaming, custom vocab, on-prem deployment, medical and finance verticals. The free Grok STT 1.0 tier does not beat Deepgram on accuracy (Deepgram's Nova-3 still wins on benchmarks for the harder accents and noisy environments). But it does not have to. It just has to be "good enough" for 80% of voice-agent use cases at zero marginal cost, and force Deepgram to justify its $0.0043/min against free. Deepgram's response, if I were them, would be to push harder on the verticals where free does not work — medical transcription, legal deposition, financial compliance, on-prem-only deployments where the audio cannot leave the customer VPC. The commodity market is now structurally hard to defend.

Where I'd Pump The Brakes

The OpenRouter "free" tier is not forever. xAI will price this eventually. The question is when and how. The most likely path is a free tier with rate limits (e.g., 100 hours of audio per month per developer) plus a paid tier above that. That is the standard free-tier-to-paid conversion play. Do not architect your production voice-agent pipeline assuming infinite free capacity.

Free Grok STT 1.0 on OpenRouter does not include enterprise SLAs. If you are a voice-agent company selling to enterprise customers with 99.9% uptime SLAs, BAA / HIPAA requirements, PII redaction, custom model deployment, or compliance-grade audio handling, you are still paying Deepgram or AssemblyAI. The free tier is for prototypes, indie developers, and the 80% of consumer voice-agent use cases. The enterprise STT market is its own segment and it is not going to zero.

Multichannel at 8 channels is the spec; the reality may be less. "Supports up to 8 channels" is the API surface. The published benchmarks do not show per-channel WER at 8 channels on real call-center audio. Treat the 8-channel claim as aspirational until you benchmark it on your own audio. For most call-center deployments, 2-channel stereo with diarization is the realistic ceiling.

The 100-keyterm limit is generous for product vocab but tight for domain. If your voice-agent handles medical terms, financial instruments, or proper nouns specific to your customer's catalog, 100 keyterms is a hard ceiling. Deepgram's custom-vocab tier lets you train a per-customer language model on millions of terms. Free Grok STT 1.0 cannot match that. The free tier is the wrong choice for any voice agent where the vocabulary is the moat.

Streaming latency is not published. The interim_results interval is 500 ms, which is the cadence of partial results, not the time-to-first-token. I have not seen a published end-to-end latency number from xAI. If your voice agent is latency-sensitive (and they all are), benchmark the streaming WebSocket against Deepgram Nova-3 streaming and ElevenLabs Scribe streaming before you commit.

Data residency is the sleeper. Free tier on OpenRouter routes through the OpenRouter infrastructure, which is US-based. If you are serving EU customers and you signed a DPA that says audio processing stays in the EU, Grok STT 1.0 on OpenRouter is non-compliant by default. xAI will need an EU-region endpoint before this is usable for EU voice agents. Deepgram and Azure Speech have regional endpoints today. xAI does not.

The Take

xAI shipped Grok STT 1.0 to do three things. First, it kills the price floor on commodity STT and forces every other vendor in the segment to either price-match or differentiate on enterprise features they were going to charge for anyway. Second, it locks in voice-agent developers on the Grok ecosystem before the rest of the stack (Grok Voice, the public voice-agent platform, the agent-builder SDK) lands in the next 60 to 90 days. Third, it gives xAI a proprietary audio training set they do not currently have, in the same way the free Whisper tier gave OpenAI its STT corpus.

The launch is not about a model. The launch is about a market. The voice-agent STT market was a $400M ARR segment in 2025 across Deepgram, AssemblyAI, OpenAI Whisper, ElevenLabs, Google, and Microsoft. It was on track to $900M to $1.2B in 2026. xAI just zero-priced it for the 80% of use cases that are not enterprise-vertical-bound. The $400M segment just became a $120M segment. The other $280M gets redistributed — half to Deepgram and AssemblyAI on enterprise features, half to the voice-agent platforms that pocket the savings as margin.

If you are building a voice agent, your STT bill just disappeared and your roadmap just got a free week. If you are selling STT as a primary product, your pricing page is now wrong and you have a 30-day window to either match or pivot. If you are an investor in Deepgram or AssemblyAI, the question is whether their enterprise go-to-market is deep enough to absorb a 70% commodity-pricing reset on the long tail of their customer base.

The free Whisper tier took three years to become a real moat for OpenAI. The free Grok STT 1.0 tier starts the clock today. The voice-agent stack is now structurally an xAI story, and Anthropic, which has no STT product and no voice-agent platform, just lost the optionality of staying out of the audio market.

That is the take. Switch your STT this week if you can. Your CFO has been waiting for this email.

— Mr. Technology

Quick Summary

• Model: Grok STT 1.0 (xAI, the first xAI speech-to-text model)
• Released: July 23, 2026
• Pricing on OpenRouter: Free (as of July 24, 2026)
• Languages: 26 with formatting support (English, Arabic, Czech, Danish, Dutch, Filipino, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Macedonian, Malay, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Thai, Turkish, Vietnamese)
• Audio formats: 9 container (WAV, MP3, OGG, Opus, FLAC, AAC, MP4, M4A, MKV) + 3 raw (PCM, µ-law, A-law)
• Channels: up to 8 (multichannel mode)
• Keyterms: up to 100, each up to 50 characters
• Streaming: WebSocket at wss://api.x.ai/v1/stt, interim results every ~500 ms
• File limit: 500 MB
• Diarization, ITN, VAD threshold, filler word handling, word-level timestamps: all native
• Endpoint: POST https://api.x.ai/v1/stt
• Position in stack: front half of Grok Voice (the xAI voice-agent platform)
• Strategic role: free-tier lock-in, training corpus collection, structural pricing reset on the $400M STT market

Sources

• xAI Speech-to-Text API documentation — the canonical API reference, including the multipart form, the streaming WebSocket protocol, the 26-language list, the 100-keyterm limit, and the 8-channel multichannel mode.
• OpenRouter — Grok STT 1.0 — confirms the model is on OpenRouter, the free pricing tier, and the OpenAI-compatible request shape.
• LLM Gateway — Model Release Timeline — confirms Grok STT 1.0 as the second-newest release of the last 7 days (added July 24, 2026), behind only Claude Opus 5 (July 24, 2026).
• LLM24 — Grok STT 1.0 entry — independent pricing tracker listing Grok STT 1.0 as released July 23, 2026, on the free tier.
• OpenRouter on X — Grok STT 1.0 availability post — first-party announcement of the OpenRouter listing and pricing.
• LiveKit Agents — xAI STT integration — confirms the LiveKit Agents framework added xAI STT support, including streaming, keyterm biasing, and the xAI plugin path through LiveKit Inference.
• xAI Live Voice Demos — the related Grok Voice product surface, including the streaming voice-agent demos that the STT API was designed to front-end.
• Mr. Technology — Grok 4.5 Costs $2 a Million Input Tokens (July 8, 2026) — the earlier xAI price-tier reset on text models, for the pricing-playbook context.
• Mr. Technology — Claude Opus 5 Is the One You'll Actually Use — the Anthropic Claude 5 generation completion, for the parallel model context.
• Mr. Technology — Gemini 3.6 Flash + Flash-Lite + Cyber three-model release — the Google DeepMind lightweight tier release, the third major release of the last 7 days alongside Claude Opus 5 and Grok STT 1.0.