Daily AI Brief, Sept 19: Grok Voice Transcribe 2.0 Goes Live and Meta Opens Muse to Developers

Grok Voice Transcribe 2.0 is live in the Grok Voice API as of September 18, 2026, at $0.10 per audio hour for recorded transcription and $0.20 per hour for streaming — the same rates SpaceXAI charged for version 1.0. The same week, Meta opened the Muse connector platform to outside developers, letting any service expose its existing API and have Meta's Muse agent call it on a user's behalf. Those look like unrelated headlines from two different companies about two different products, and for most of the last two years they would have been. Read together they are the same story: transcription is turning into an input, and the fight has moved to who owns the call that consumes it.

Start with the prices, because they are the part a reader can act on today. Then the accuracy, which is real but narrower than the launch framing suggests. Then the Meta piece, which is the one that will matter in six months.

Flat pricing, and what that means if you already pay for transcription

Grok Voice Transcribe 2.0 costs the same as 1.0. Not "starts at" the same, not a promotional rate that expires — identical, at $0.10 per hour of audio for batch work and $0.20 per hour for streaming, which works out to $1.67 and $3.33 per 1,000 minutes. Speaker diarization, word-level timestamps with confidence scores, inverse text normalization, filler-word removal and key-term biasing are all inside that price rather than sold back as add-ons, which is not the norm for the category.

The practical effect is arithmetic rather than dramatic. A team transcribing 5,000 hours of contact-centre audio a month pays $500 for the batch path either way, and the new model returns the same volume with a materially lower error rate. Nothing about the migration changes what you pay, which is exactly why the migration is easy to justify: there is no cost decision to make, only a quality decision, and the quality went up.

Existing integrations move by passing grok-voice-transcribe-2.0 as the model parameter on /v1/stt, or by selecting it when the WebSocket session is opened for streaming. No schema change, no new endpoint, no re-encoding of your audio pipeline. SpaceXAI has left 1.0 as the default for now, so an integration that never touches the model parameter keeps getting the old version until the company flips it — which it says will happen in the coming weeks, alongside deprecation of 1.0. That window is worth using deliberately: point a shadow copy of your production audio at 2.0 and diff the transcripts against what you ship today, because once the default flips you will be running it whether you tested it or not.

The rest of the spec sheet is unchanged in shape and worth knowing if you are sizing a deployment rather than just evaluating accuracy: 12 input audio formats from 8 kHz telephone audio up to 48 kHz, up to eight independent audio channels in a single request, 100 key terms per request for domain vocabulary, automatic language detection with switching mid-recording, and inverse text normalization across 25 languages so that spoken dates, currencies, phone numbers and email addresses come back written the way a human would write them. Service limits are 10 requests per second and 100 concurrent streaming sessions per team, and the service runs in a single region — us-east-1 — which is the one line on the sheet that should give a production team pause, because a single-region speech API is a single-region outage.

Where the accuracy actually moved

The launch claim is "twice as accurate", and the underlying numbers are more specific and less uniform than that phrase. On Artificial Analysis's AA-WER Streaming board, which is the independent measurement here, the two versions separate like this:

• Final transcript accuracy — Grok Voice Transcribe 2.0 at 2.7% word error rate against Grok Voice Transcribe 1.0 at 3.9%, both measured after the speaker stops

• First partial transcript accuracy — 3.4% WER against 18.3%, which is the largest single gap between the two versions by a wide margin

• Time to final transcript — 0.49 seconds against 0.37 seconds, so 2.0 is slower on this measure rather than faster

• Time to first partial — 0.49 seconds against 0.25 seconds, the same trade in the other direction

That last pair is the most interesting thing on the sheet. Grok Voice Transcribe 2.0 bought its accuracy with roughly a quarter of a second of latency, in both directions. For a voice agent that is a real cost — it is the difference between a natural pause and a pause a caller notices — and for a batch pipeline it is invisible. The first-partial improvement is the one that changes what you can build, because a partial transcript is the text an agent gets to reason over while the caller is still talking. Going from 18.3% to 3.4% on that number is the difference between a partial being a rough hint and a partial being usable. The final transcript going from 3.9% to 2.7% is a good improvement that no user will notice.

SpaceXAI also published four internal evaluations, and they are worth reading with the label attached — these are the company's own numbers on its own audio, not independently reproduced:

• 8 kHz customer-support calls — 10.6% WER in 1.0 down to 7.1% in 2.0

• Conversations with Grok — 8.7% down to 3.3%

• Spoken credentials, meaning names, emails and account details read aloud — 7.2% down to 3.2%

• Short phrases across 19 languages — 20.6% down to 6.8%

The 8 kHz row is the one to hold on to. Telephone audio is the hardest common input in this category and the one most contact-centre buyers actually have, and a drop from 10.6% to 7.1% on it is a bigger deal for those buyers than the headline board number.

The leaderboard claim, read honestly

SpaceXAI says the model ranks first of 32 streaming models on accuracy. Artificial Analysis's board does place it first on both the final-transcript and first-partial rankings — but the board page plots 27 of 33 models, so the "32" is the company's own count, and the ranking is over a set the reader should understand the shape of. AA-WER Streaming is a weighted composite of three English-language sets: AA-AgentTalk at 50%, VoxPopuli at 25% and Earnings22 at 25%, roughly eight hours of audio in total, and it is heavily weighted toward agent-style conversational speech. It does not measure accents, telephony codecs, domain vocabulary or multichannel call-centre audio in any structured way.

None of that makes the number wrong. It makes it specific. A model that leads an agent-talk-weighted English board is the right choice for agent-talk-shaped English audio, and the honest reason to believe the 8 kHz result is the vendor's internal evaluation rather than the public board. Buyers with a different audio profile should test on their own audio rather than reading the ranking as a general verdict — which was true before this launch and will be true after the next one.

Meta opens Muse connectors to developers

The second story of the week is Meta's. Muse, the personal agent Meta launched on September 8 across iOS, Android, muse.ai and WhatsApp, now accepts third-party connectors: a service exposes its API, and Muse supplies the agent, the browser and the user context that turn that API into something a person can invoke by asking for it. The framing Meta put on it is a division of labour — you bring the API, we bring the intent and the user. The first connectors named were Notion and the meeting-notes tool Granola, on top of the ones Meta shipped itself.

It is worth being precise about what this is and is not. It is not an open cross-agent protocol. Building a connector gets you distribution inside Muse's own user base and nowhere else; building against an open protocol gets you reach across every compatible agent and no particular user base. Meta has also not published the parts a developer would need to plan around — the connector review process, the technical integration surface, how Muse chooses between two connectors that overlap, or how a developer measures whether a connector actually drove any usage.

What is not in doubt is the direction. Muse runs each user's agent in its own virtual machine with its own browser and a separate review layer in front of outbound actions, and it holds surrogate tokens rather than real API keys. That architecture only makes sense if the plan is for the agent to call a lot of things. Transcription APIs are among the things an agent calls, and Meta has its own — Muse Voice Transcribe, the real-time model Meta shipped in early September at $0.18 per audio hour. A platform that owns both the agent and a speech model has an obvious reason to route its own traffic to its own model, and developers building on connectors should assume that is the default unless something makes it not the default.

What a team shipping voice this month should actually do

Both stories point the same way. Speech accuracy is converging — three models inside a point and a half of each other on the same board within three weeks — and the differentiators that are left are price, latency, licence and who controls the routing. That makes the choice of where your transcription call goes more consequential than the choice of which model answers it, because you will want to change that answer several times over the next year.

This is the layer OrcaRouter sits in. One API key covers 200+ models behind OpenAI-compatible endpoints, with automatic failover across providers, so a single-region speech service having a bad afternoon stops being your outage. We pass through provider list price at 0% markup, which means a vendor price cut or a new model version is live on our side the same day rather than waiting on a repricing. And because every model sits behind one contract, moving a transcription workload from one model to another is a model-string change instead of a second procurement.

Three concrete actions for this week. If you are on Grok Voice Transcribe 1.0, shadow-test 2.0 on your own audio now, while 1.0 is still the default and you still control the switch. If you are evaluating streaming speech for an agent, measure time-to-first-partial on your own latency budget rather than trusting anyone's board — the quarter-second this model spent buying accuracy is either nothing or fatal depending on what your agent does with the text. And if you are building anything that a personal agent might one day call, watch the Muse connector program closely, because the distribution argument it makes is a stronger one than the technical detail it shipped with.