Meta folds diarization and endpointing into the transcript stream for $0.18 an hour

The unit of purchase here is a token stream. Audio arrives in 80 ms chunks, 12.5 per second, each compressed into one soft token ^[11]. At every chunk the model picks one of two actions: emit a text token, or emit a "next audio" placeholder that the runtime swaps for the following chunk ^[12]. When the audio stops, an "empty audio" token tells the model nothing more is coming, and it flushes whatever text it is still holding ^[13]. Speaker turns ride that same sequence. A `<|start_of_turn|>` token marks a possible change, a token such as `<|speaker_A|>` assigns the turn, and separate tokens mark start and end of speech, all trained alongside transcription rather than handed to a downstream speaker classifier ^[16].

That gives you something to budget with. 12.5 chunks a second across an hour is 45,000 soft tokens ^[1], and $0.18 for that hour works out at $4 per million audio tokens ^[2]. Streaming and batch carry the same rate, and diarization is not a separate line ^[2].

The reward design is the part worth copying. Meta multiplies the word-error-rate reward by the delay reward instead of adding them ^[15]. Addition lets a model buy accuracy with unlimited stalling, because a strong WER term covers a weak delay term. Multiplication makes a near-zero score on either factor collapse the whole product. That is what produces adaptive delay, where the model spends more audio on an ambiguous word and commits early on an easy one ^[14]. What the launch material does not give you is a millisecond figure; the only published timing value is the 80 ms chunk period ^[7]. Barge-in budgets are a distribution here, not a constant, and you will have to measure it yourself.

For the accuracy number to describe your workload, your audio has to be English, streamed, and close enough to whatever mix Artificial Analysis assembled, because that benchmark covers English only ^[17]. Meta says 25 of more than 70 trained languages were extensively validated for this release ^[5], so anything outside those 25 is your eval problem. The margin over the next model on the list is 0.3 points ^[6], in a field where OpenAI, Google, xAI and Alibaba all shipped streaming models within weeks of each other, according to The New Stack ^[21].

The bundled feature is also the weak one. Speaker error at 17.5% is roughly 5.6 times the word error rate ^[5]. If your interface prints a name next to a quote, 17.5% is what the user meets.

On price, the bundle undercuts the $0.24 Speechmatics lists for Standard real-time transcription by 25% ^[9][3], while Soniox posts about $0.12 an hour with diarization, language identification and formatting included, a third under Meta ^[8][4]. The posted floor belongs to Soniox. What moved is the default assembly: one vendor billing one rate for work that used to need three integrations. Speechmatics will label 50 speakers by default and 100 on request, five times Meta's stated 20-plus ^[9][8], and if a hundred people are talking at once, speaker labels are not your first problem. Amazon Transcribe documents 30 ^[10]. Meta is not releasing these weights, a spokesperson told The New Stack ^[19], so a consolidated pipeline is a hosted dependency with no local build behind it.