The cheapest speech-to-text you can buy is now a one-line change. As of July 23, xAI's Grok STT 1.0 is listed on OpenRouter, which means any agent already routing through OpenRouter can add transcription with a model string and its existing key — no new xAI account, no new SDK. The batch rate is $0.10 an hour. That undercuts every incumbent, and for a founder shipping a voice feature on a budget, that is the headline.

It is also the only number in this comparison you can actually trust. So before you rip out Deepgram, it's worth separating what's verified from what's marketing.

What's real: the price#

Grok STT's $0.10/hour batch and $0.20/hour streaming are public and identical whether you call xAI directly or route through OpenRouter. Line them up against the list rates and the gap is real, not rounding:

At a thousand hours a month, that's the difference between a $100 bill and a $260 one. The feature set is table stakes for the category: 25+ languages, word-level timestamps, speaker diarization, multichannel, and inverse text normalization, with a REST batch endpoint and a WebSocket for streaming. Nothing missing that would disqualify it.

What's not real yet: the accuracy chart#

xAI's launch numbers are excellent — a 6.9% general word error rate, and on phone-call entity recognition (names, account numbers, dates) a 5.0% error rate against ElevenLabs at 12.0%, Deepgram at 13.5%, and AssemblyAI at 21.3%. If those held on your audio, this would be the easiest switch of the year.

They are also, every one of them, xAI's own benchmarks, run by xAI, unverified by anyone else. No vendor has ever published a benchmark that made it look bad. And xAI's own chart quietly undercuts the story: on video and podcast audio, Grok ties ElevenLabs at 2.4% — so the dramatic lead only appears on the noisy, entity-dense phone-call set that happens to flatter it most.

Treat a vendor's self-reported accuracy as a marketing floor, not a measured ceiling. The only accuracy number that matters is the one you get running your own audio through it.

This is not a reason to dismiss it — Deepgram and AssemblyAI earned their reputations on independent benchmarks, and Grok STT hasn't been through that gauntlet yet. It's a reason to A/B it on a real slice of your traffic before you trust a transcript to it.

The quieter catch: one provider, no failover#

There's a second line in the OpenRouter listing that matters more than it looks: Grok STT 1.0 is hosted by a single provider, forwarded directly, with no redundancy layer. Most mature models on OpenRouter fan out across several backends, so a provider outage is invisible to you. Here it isn't — if that one provider hiccups, your transcription is down. The whole appeal of routing through OpenRouter is resilience, and this model doesn't have it yet.

That makes the architecture obvious: if you adopt Grok STT, put a fallback behind it the same way you would a cheap model with a frontier backstop — Grok STT primary for cost, Deepgram or AssemblyAI as the catch. OpenRouter's own routing already makes swapping models a config change, not a rewrite, so the fallback is cheap to wire.

The actual decision#

Don't read this as "who's most accurate" — you can't answer that from vendor charts. Read it as where each one belongs:

The voice-agent transcription landscape we mapped in June just got a fourth serious entrant that competes on the one axis founders feel first — the bill. Whether it competes on the axis that actually matters — the transcript — is a test you now have to run yourself, at ten cents an hour.