Grok Voice Transcribe 2.0: 2x accuracy at the same $0.10/hr
xAI's STT tops the streaming-accuracy leaderboard with diarization free — and Claude Code stops billing auto-mode overhead and reads AGENTS.md.
Copy markdown
2x accuracy, same price, diarization free
Grok Voice Transcribe 2.0 is roughly twice as accurate as v1 at the same rate — voice-command word error rate fell from 20.6% to 6.8%, and streaming now lands at 2.7% just 0.49s after speech ends. Batch stays $0.10/hr, streaming $0.20/hr, with speaker diarization, word-level timestamps, and key-term biasing (up to 100 terms) all included. Enough to wire into a voice agent tonight.
It tops 32 streaming STT models
xAI says v2.0 ranks first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard, ahead of Gemini 3.5 Transcribe, Deepgram Nova-3, and Whisper Large v3 on internal telephony sets. It handles up to 8 channels, files to 500 MB, and switches languages mid-recording in a single pass. It becomes the Speech-to-Text API default soon — pin grok-voice-transcribe-1.0 to stay on the old model.
Claude Code stops billing auto-mode overhead
Claude Code 2.1.278 makes auto mode default to a server-side classifier that doesn't charge for classifier overhead — for API, Enterprise, Bedrock, Vertex, Foundry, and gateway users. A new /status row shows whether yours runs server-side; opt out with CLAUDE_CODE_AUTO_MODE_SERVER=0. Small line item, real savings if you run auto mode all day.
Claude Code now reads AGENTS.md
Version 2.1.277 adds AGENTS.md support: in a repo with no CLAUDE.md, Claude Code reads AGENTS.md instead (toggle under Project instructions in /config). If you already keep an AGENTS.md for Codex or other agents, that's one less file to maintain. Not yet available on Bedrock, Vertex, or Foundry.
Codex CLI adds experimental /voice
OpenAI's Codex CLI v0.155.0 adds an experimental /voice mode with live transcripts and mic controls (behind /experimental), plus live reasoning summaries in the status row and Touch ID verification for MCP requests on supported Macs. Talk to your coding agent instead of typing.