Grok Voice Transcribe 2.0 halves errors at the same price
xAI's speech-to-text now tops 32 streaming models on accuracy at $0.10 an audio hour — as Meta ships a Mac agent and Linkup opens a tiny sparse embedder.
Copy markdown
Half the errors, no price change
Grok Voice Transcribe 2.0 is twice as accurate as 1.0 at the same rate and ranks first among 32 streaming models on Artificial Analysis's leaderboard — its multilingual short-phrase error rate falls from 20.6% to 6.8%. Batch runs $0.10 per audio hour, streaming $0.20, so accurate transcription just got cheaper per correct word.
Diarization, timestamps, and term biasing come included
Every call ships word-level timestamps, speaker diarization across up to 8 channels, key-term biasing (100 terms per request), and smart turn detection, plus auto language detection that follows mid-recording switches in one pass. It's live in the Speech-to-Text API and becomes the default — pin `grok-voice-transcribe-1.0` if you need to stay put.
Elsewhere: Meta drops a task-running agent onto your Mac
Muse for Mac is a personal agent that acts across your files, mail, messages, calendar, and notes on macOS — Meta's move into desktop computer-use agents. Builders get separate hooks through Muse Code and the Muse Spark model API.
Elsewhere: Linkup opens a 149M sparse embedder for cheap search
SPARSEUP (Apache 2.0, on Hugging Face) emits vocabulary-weighted sparse vectors that drop straight into an inverted index — 56.4 nDCG@10 on BEIR-13, the best public sparse encoder under 150M params, at ~380µs per query single-threaded. Human-readable weights, no dense-vector database required.