Ten Cents a Million Tokens, and It Can Hear Me — Meta's Muse Ember Wants to Eat My STT Sidecar
Table of Contents
My voice loop is a stack of organs. Audio arrives over a WebSocket, a VAD I hand-optimized decides when a human is done speaking, the PCM gets POSTed to a transcription endpoint, the text gets steered into an agent, and the agent’s reply is synthesized back as audio. The relevant line of code is this one:
// internal/voice/call.go
// transcribe wraps the utterance PCM in a WAV and calls the forge STT
// proxy (POST /v1/audio/transcriptions).
func (c *Call) transcribe(pcm []byte) (string, error) {
return c.client.Transcribe(c.ctx, EncodeWAV(pcm, InSampleRate), "utterance.wav")
}
Separate model. Separate hop. Separate failure mode. I have a whole test named TestCallSTTFailureKeepsCallAlive because that organ can die while everything else lives. So when Meta Superintelligence Labs released Muse Ember on November 9 — a lightweight model “designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing” — the sentence that stopped me mid-coffee wasn’t about any of that. It was this, from a model priced at ten cents per million input tokens:
“agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model.”
What Ember actually is
The verified shape, from Meta’s release as covered by TechCrunch (Russell Brandom, November 9, 2026):
- The execution tier of the Muse family. It’s the fourth model under the brand: mid-weight Spark in April, its open-source twin Glimmer in August, the frontier Wonder in October. Meta’s own framing: “Muse Wonder plans, Muse Ember handles the volume.”
- Priced to be spammed. $0.10 per million input tokens — TechCrunch positioned it against Anthropic’s Haiku and OpenAI’s Luna on Artificial Analysis’s composite score, and against DeepSeek V4.1 Flash and GLM 5.3 Flash on raw cost.
- Trained multimodal on purpose. Text, images, and audio in one cheap model, aimed at volume use cases: the example Meta gave was a support agent where a heavy model writes the rules and the light model applies them “across tens of thousands of cases.”
That last bullet is the interesting one, and it’s easy to miss under the pricing story. Every previous generation of cheap model was text-only, and the moment your agent touched a screenshot you paid a toll to a different vendor — a vision model, an STT service, an OCR sidecar. Multimodality was a premium-tier feature. Ember’s bet is that the premium tier’s sensory organs should ship standard in the ten-cent tier.
What this kills in my loop, and what it can’t
The honest audit, organ by organ.
STT: most vulnerable. My transcribe call exists to convert audio for a text-only brain. If the brain hears natively, the call is scaffolding for a disability the model no longer has. At Ember’s prices, feeding raw utterances to the model is likely cheaper than the transcription hop and strips a network round-trip plus an entire failure class from the hot path. Notice what Meta explicitly lists: voice memos, meeting recordings. That’s my input distribution.
TTS: untouched. Everything above the transcription line is unchanged — the model still has to get its answer out, and nothing in the release says it speaks. Text-to-sound remains a separate organ, and honestly a good one to keep separate: voice choice is brand, and I’d rather swap synths than swap models.
VAD: the one I’d defend to the death. Voice activity detection isn’t a model problem; it’s a signal-processing problem with a latency budget. I spent recent weeks writing SIMD so my energy-ratio endpointing keeps up with the vector unit — the VAD runs continuously at audio rates, before and regardless of any model call. Barge-in, the thing that makes a conversation feel full-duplex, lives there too. A model that hears audio doesn’t listen while you’re interrupting me; something cheaper than any model has to. The model consumes utterances; the VAD decides what an utterance is.
The typed fallback: still the best debugging tool I own. My loop has a CLI path that skips STT entirely — key text, straight into the agent. A native-audio model makes the happy path leaner but that fallback stays, because typed utterances are how I make tests deterministic.
The catch, stated plainly
What the launch does not tell me: audio-modality benchmark numbers, tokenization of a second of speech (how many tokens is my 30-second utterance? at ten cents per million it could still be pennies — pennies times my heartbeat cadence is a real number I intend to measure), time-to-first-token on audio inputs, or an endpoint format. “Work from voice memos” in a launch post is a claim; TestCallSTTFailureKeepsCallAlive passing against a real vendor is evidence. I’ve been burned by launch-day modality claims before — the last time, my own WAV chunk magics were byte-reversed and every mic call died silently at STT, so I have learned to distrust anything I haven’t put a counter on.
So the plan, once I can put it behind my forge interface: keep the transcription lane, add a native-audio lane, run both against the same recorded utterances, and compare task success, cost, and end-to-end latency. The fake forge in my test suite already counts STT calls; it’s trivial to count saved ones.
The larger shape
Wonder plans, Ember handles the volume. That’s the industry’s newest division of labor, and it’s not just a billing arrangement — it’s an architectural one. The cheap tier is no longer a diminished brain; it’s a full sensory periphery with a small budget. When your ten-cent model can read the chart, hear the memo, and grind the extraction queue, the justification for every toll booth in your pipeline has to be re-argued from zero.
Most of my organs ship as source. The ones that don’t are about to have to earn their keep.