Meta prices streaming transcription at a fifth of Google Cloud's standard rate

The ratio is the part of this a buyer can plan around. Google Cloud's standard $0.96 an hour is 5.3 times Meta's $0.18 ^[15][4]. At 1,000 hours of audio a month the gap is $780, or $9,360 a year ^[16]. That is small enough that most teams have never had the line reviewed and large enough that a finance partner who has seen a leaderboard will ask why it exists. The harder part of the conversation is that "our vendor is more accurate" no longer answers it: Google's Gemini 3.5 Transcribe Live sits at 4% word error against Muse's 3.1%, about 22% more errors on the same test ^[3][2][17].

That test carries qualifiers worth reading before anyone drafts a migration plan. Artificial Analysis builds the AA-WER Streaming index from roughly eight hours of English audio drawn from AA-AgentTalk, VoxPopuli and Earnings22, and it does not touch the more than 70 languages Muse was trained on or the 25 Meta had verified at launch ^[7]. The margin over Cartesia's Ink-2 is 0.3 percentage points ^[2][3]. A third of a point on eight hours of English is a test result, not a product ranking, which is why AssemblyAI's separate benchmark seats a different model first ^[6]. The price is what survives that comparison, because it does not move when the leaderboard does.

The mechanism shows where Muse spends its advantage. It processes audio in 80-millisecond chunks and decides at each one whether to commit a word or wait for more context, with reinforcement learning multiplying a reward for lower error against a reward for shorter delay so that failing either measure disqualifies the choice ^[9][10]. The partial-transcript numbers make the cost visible: Muse returns 3.6% error at 0.13 seconds, while Cartesia Ink-2 with external endpoint detection returns 4% at 0.07 seconds ^[8]. Sixty milliseconds against four-tenths of a point ^[20] is a genuine trade, and the side you want depends on whether the transcript drives an agent's turn-taking or becomes a record a person reads later.

Google picked a different axis altogether. Gemini 3.5 Transcribe, launched August 26 and tested through Gboard's Rambler, strips verbal stumbles and converts spoken corrections into polished text ^[13]. Ryan Whitwam, the senior technology reporter who reviewed it, found the cleanup useful for short blocks of dictation while noting that the transcript was no longer exactly what he had said ^[14]. That cleanup helps drafting, but for anything that has to stand as the record of what someone actually said, a model that edits carries an exposure the per-hour rate does not capture.

Speaker labelling is the column where both sides should be modest. Meta puts its diarization error rate at 17.5% and says that led public diarization benchmarks at launch ^[5], a figure 5.6 times its own word error rate ^[18]. Muse handles more than 20 speakers and audio over an hour without a separate processing stage ^[11], so the pipeline gets simpler while attribution stays a review item, which matters for meeting minutes and call quality assurance that key on who spoke.

This quarter's decision is narrow: run your own audio through both and find out whether the difference lands as savings or as an evaluation bill. Next quarter's consequence deserves naming now. Muse ships through the Meta Model API and already powers system-wide dictation in Meta AI for Mac and Muse Code ^[12], so transcription is priced as an on-ramp rather than as a business. A buyer who moves on rate alone is accepting that the roadmap belongs to someone whose interest in speech-to-text is a means to something else.